The GEO/AEO playbook: how to become the cited answer in ChatGPT, Perplexity, and Google AI Overviews.

Most teams approach generative engine optimization the way they approached SEO in 2012: publish more pages, add a schema plugin, and wait. That worked when the search engine ranked ten blue links and the user did the synthesis. It does not work when the engine writes the answer and cites three sources.

In an answer engine, you are not competing for a position. You are competing to be the evidence behind a sentence. That changes what you publish, how you structure it, and how you prove it worked.

This is the operating model we use across GEO and AEO engagements: five workstreams, one measurement layer, and a 90-day cadence. It is deliberately unglamorous. Most of the lift comes from making the business legible to a retrieval system.

What actually changed in 2026

Three shifts matter more than the rest.

First, retrieval got better and less predictable. ChatGPT Search, Perplexity, Google AI Overviews, Claude, and Gemini all combine a model with a live retrieval layer. They do not share an index. A page that ranks well in Google may never be retrieved by Perplexity, and a page with no traditional rankings can become a primary citation if it answers a specific query cleanly.

Second, entity confidence beats keyword coverage. When a model summarizes a company, it checks whether the facts agree across sources: the site, LinkedIn, Crunchbase, Wikidata, review platforms, press coverage, and structured data. Contradictions reduce confidence. Low confidence produces vague answers or a competitor recommendation.

Third, the click is no longer the default. A growing share of high-intent research ends inside the answer. That does not make the work less commercial. It makes being the cited source the commercial outcome, because the buyer arrives pre-qualified by a sentence you influenced.

The practical implication: GEO is not a content channel bolted onto SEO. It is a data-consistency and retrieval program with content as the interface.

Start with a prompt map, not a keyword list

Keyword tools tell you what people typed into Google. They do not tell you what a buyer asks an assistant when they are three weeks into a vendor evaluation. You need both, but the prompt map comes first.

We build the map in four steps:

1. Harvest real language. Pull questions from sales calls, support tickets, onboarding forms, community threads, and the "People also ask" layer. Aim for 60–120 raw questions.
2. Cluster by decision stage. Group prompts into problem-aware, solution-aware, vendor-aware, and decision. A prompt can appear in more than one cluster if the intent changes.
3. Test the current answer. Run every prompt across the engines you care about, in the markets you sell into, with a clean session. Record who is cited, what claim is made about you, and whether the answer is accurate.
4. Assign one page per cluster. Not one page per prompt. The page should be able to answer the cluster's core question in the first 120 words, then support it with evidence.

The output is a working document, not a report. We keep it in a spreadsheet with columns for engine, locale, date, cited sources, brand mention, sentiment, and accuracy. That sheet becomes the baseline for every later claim about progress.

One warning: do not optimize for prompts you would never sell against. A citation in "best growth agencies for enterprise fintech" is worth more than fifty citations in "what is growth marketing". Volume feels good; relevance pays.

Build the entity before the page

An answer engine needs to know who you are before it can decide whether to cite you. That sounds abstract until you see the failure mode: the model describes an outdated product, lists a discontinued pricing tier, or recommends a competitor because your category is ambiguous.

We fix this with a canonical fact set — a single page or document that defines the business in machine-readable terms:

• Legal and trading names, including former names and common misspellings.
• Category in the buyer's language, plus the adjacent categories you do not want to be confused with.
• Products and services with plain descriptions, pricing model, and who they are for.
• Locations and markets served, with currency and language coverage.
• Proof points that are stable enough to repeat: certifications, integrations, security posture, notable clients where permitted.
• Canonical URLs for every one of those facts.

Then we reconcile the outside world against that set. Wikidata, LinkedIn, Crunchbase, G2, Clutch, industry directories, and press coverage all get checked. Where a fact is wrong, we correct it at the source. Where a source cannot be corrected, we publish a stronger canonical page and link to it consistently.

Schema is a translator, not a strategy

Structured data helps a machine parse what is already on the page. It does not create authority. We use a focused set: Organization, WebSite, Service, Product or SoftwareApplication where relevant, FAQPage for genuine question-and-answer blocks, Article for editorial work, and BreadcrumbList for hierarchy.

The common mistake is marking up claims that the page does not visibly support. If the schema says you serve 40 countries and the page lists three, you have created a contradiction, not a signal. Mark up what a human can verify.

Write citation-ready answers

Retrieval systems favor passages that can stand alone. A paragraph that begins with "As we discussed above" is useless when extracted. A paragraph that names the entity, states the claim, and supplies the evidence can be quoted directly.

We structure every answer page the same way:

1. A direct answer in the first 80–120 words. Name the entity, the outcome, and the condition under which the answer holds.
2. The mechanism. Explain how it works in enough detail that a model can distinguish you from a generic vendor.
3. Original evidence. A benchmark, a teardown, a dataset, a screenshot, a customer result with a method note. This is what makes you quotable rather than paraphrased.
4. Boundaries. When the answer does not apply. Counterintuitively, this increases citation quality because it signals expertise and reduces mismatched retrieval.
5. Related entities. Link to the service, the proof, and the next step using descriptive anchor text.

Headings should read like the questions buyers actually ask. "How much does a GEO engagement cost?" outperforms "Investment" because the retrieval layer can match the passage to the prompt without inference.

In practice, this often means replacing a single 3,000-word "ultimate guide" with twelve focused answer pages: integrations, security review, migration path, pricing model, competitor comparisons, and implementation timeline. The new pages are shorter and easier to retrieve. Whether that improves citation share depends on the category, the competition, and how the engines source answers — so it is measured, not assumed.

Make the site retrievable

You can do everything above and still lose because the crawler cannot read the page. Retrieval crawlers are less patient than Googlebot. They often do not execute heavy client-side JavaScript, and they may time out before your hydration completes.

The technical checklist we run before content work:

• Server-render the answer. The first screen of meaningful copy should exist in the HTML response, not after a client-side fetch.
• Keep response times low. Target sub-650ms TTFB at the edge. Slow origins get crawled less.
• Allow the relevant agents. Review robots.txt for GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot, and Google's crawlers. Blocking training crawlers is a business decision; blocking search crawlers is usually a visibility mistake.
• Publish a useful /llms.txt. It is not a ranking factor, but it is a clean map of your canonical pages. See the implementation guide.
• Use stable URLs and canonical tags. Duplicate variants split retrieval signals.
• Compress images and avoid layout shift. Retrieval quality and page experience share the same delivery layer; fix it once.

We verify access with a server-side fetch that spoofs the agent string, not with a browser. If the HTML response does not contain the answer, the page is not ready.

Measure citation share, not vibes

The fastest way to kill a GEO program is to report something the finance team cannot tie to revenue. We use three layers.

1. Citation share and share of answer

For each prompt cluster, record whether you are cited, how prominently, and whether the surrounding answer is accurate. Citation share is the percentage of tracked prompts where you appear. Share of answer is a weighted version that gives more credit to prompts where you are the primary recommendation rather than a footnote.

Run the panel weekly in a clean session, in each target locale. Store the raw answer text. Engines are non-deterministic; a single snapshot is not evidence.

2. AI referral behavior

Segment analytics for referrals from chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com, and claude.ai. Measure landing page, engagement, and conversion rate against your site average. In our experience, AI-referred sessions convert well above average because the buyer has already read a synthesized answer.

3. Assisted pipeline

Add a self-reported attribution question to demo and contact forms: "How did you first hear about us?" It is imperfect, but it catches the research that never produces a click. Match those responses to CRM opportunities and report AI-influenced pipeline separately from last-click.

The 90-day operating cadence

GEO is not a launch. It is a loop. We run it in three 30-day blocks.

Days 1–30: establish the baseline

• Build the prompt map and run the first three panels.
• Audit entity consistency across the site and the top ten third-party sources.
• Fix crawler access, rendering, and the worst page-speed regressions.
• Ship the canonical fact set and a first version of /llms.txt.

Days 31–60: ship the answer layer

• Publish or rewrite the six highest-leverage answer pages.
• Add original evidence to each: a benchmark, a teardown, or a customer result.
• Rewrite internal links around topic clusters rather than legacy navigation.
• Correct the third-party facts that still contradict the canonical set.

Days 61–90: compound and prune

• Re-run the prompt panel and compare citation share by cluster.
• Double down on pages that earned citations; merge or prune pages that did not.
• Add comparison and integration pages where vendor-aware prompts are underserved.
• Connect AI referral and self-reported attribution to CRM pipeline.

After day 90, the cadence continues monthly: one new evidence asset, one entity correction pass, one prompt panel, one pipeline review. The work is small but constant. That is what compounding looks like in practice.

Where GEO programs fail

We have audited enough programs to see the same failure modes repeatedly.

• Publishing volume without retrieval structure. Fifty posts that never state a clear answer will lose to five pages that do.
• Schema without substance. Markup cannot rescue a page that does not say anything verifiable.
• Optimizing for the wrong prompts. High-volume, low-intent questions produce citations that never influence a purchase.
• Ignoring third-party facts. Your site can be perfect while Crunchbase describes a product you retired two years ago.
• Reporting screenshots instead of panels. One lucky ChatGPT answer is not a trend. A repeatable panel with dates and raw text is.
• Treating GEO as a marketing-only project. Product, legal, security, and founder-led content all supply facts the model needs.

The good news is that the bar is still low. Most competitors are publishing generic advice and hoping. A team willing to define its entity, answer real questions, and measure citation share honestly can become the default answer in a category faster than it could rank for the same terms organically.

If you want to see where you stand, start by running a small prompt panel yourself, then read the llms.txt implementation guide for the technical layer. If you would rather have us build the panel and the roadmap, request a proposal.