SEO, GEO, and AEO: How Companies Get Found in AI Search

What changed: the answer replaced the list

For twenty years, search visibility meant a position in a list. A buyer typed a few words, scanned ten links, and chose. Every optimization technique was aimed at that list.

Generative engines behave differently. A buyer asks a full question — who can retrofit an obsolete PLC line near Chicago without replacing the machine — and receives a synthesized answer naming a handful of companies, with a few citations attached. There is no page two. Either your company is in the answer, or it is not part of the decision.

This does not make SEO obsolete. Generative engines retrieve from the same crawled and indexed web, so pages that rank well are usually the pages that get retrieved. What changes is the finish line: no longer a rank, but a sentence about your company that a model is willing to generate and attribute.

SEO, GEO, and AEO defined precisely

The three acronyms overlap heavily and get used loosely by vendors. The distinction that actually matters is the unit of success.

• SEO — search engine optimization. Target: ranked positions in a classic results page. Levers: crawlability, indexation, relevance, page experience, internal linking, and authority signals.
• AEO — answer engine optimization. Target: the direct answer, including featured snippets, People Also Ask, voice responses, and AI Overviews. Levers: question-shaped headings, answer-first paragraphs, FAQ structured data, and unambiguous claims.
• GEO — generative engine optimization. Target: being retrieved, quoted, and cited inside a generated answer. Levers: machine-readable entity facts, self-contained statements that survive being pulled out of context, and corroboration across sources a model can check.

One system, not three projects

Roughly four fifths of the work serves all three at once: HTML a crawler can read, structured data that validates, facts that agree with each other, and writing that answers real questions. The remaining fifth is where GEO and AEO ask for things classic SEO never did — a company published as data, and prose engineered to be quotable in fragments.

That is why this belongs to engineering as much as to marketing. Rendering pipelines, routing, schema generation, and build-time validation are not content decisions, and a content team cannot fix them.

Why companies disappear from AI answers

Five failure modes account for most cases. All five are diagnosable in an afternoon.

1. The site renders only in JavaScript. GPTBot, ClaudeBot, PerplexityBot, and most AI crawlers do not run JavaScript. A React, Angular, or Vue application that assembles content in the browser can return an empty shell to them while looking perfect to you. Open view-source rather than the inspector and search for a sentence from your page; if it is missing there, you are missing from the answer.
2. The company has no machine-verifiable identity. When no structured data states the legal name, location, services, and contact details, a model must infer them from prose — and grounded systems decline to assert facts they cannot confirm.
3. The content is marketing prose instead of answers. A slogan such as 'innovative solutions for modern manufacturing' matches no question anyone asks. Retrieval works on similarity to a real question, so a page that never states the question cannot be similar to it.
4. The facts contradict each other. A different address in the footer than in the schema, three variants of the company name, a capability claimed on one page and denied on another. Contradiction is the fastest route out of an answer, because the safest thing a grounded system can do with conflicting evidence is cite somebody else.
5. Nothing on the page is quotable in isolation. Retrieval pulls fragments, not pages. A sentence that needs the previous three paragraphs to make sense loses to one that stands on its own.

Layer 1 — make the page readable to machines

This layer decides whether anything else matters, and it is the layer most marketing agencies cannot deliver, because it is software work rather than copywriting.

• Server-side rendering or build-time prerendering, so every public route returns complete HTML with content, meta tags, and JSON-LD in the raw response.
• One canonical URL per page, correct status codes, and no soft 404s. A route that returns 200 with an empty body teaches crawlers that your pages are empty.
• Parity between what a browser shows and what a crawler receives. Serving different content to bots is cloaking and is penalized; the goal is identical content delivered without requiring JavaScript.
• A robots.txt that makes a deliberate decision about each AI crawler — GPTBot, ClaudeBot, PerplexityBot, Google-Extended, CCBot — rather than blocking them by copied default.
• Fast responses and lean pages. Crawl budget is real and retrieval systems time out.

How to verify layer 1 yourself

Request your own page with a crawler user agent and read what comes back: curl -A GPTBot on the URL you care about, then search that text for the sentence you most want quoted. If it is not there, no content strategy will rescue it.

Repeat the test on pages produced by your CMS or API. They frequently render differently from hand-built pages, and that gap is the usual reason new content ships invisible while the older pages test fine.

Layer 2 — publish your company as data

A generative engine asserts what it can ground. Give it grounding in formats it already parses.

• Organization and LocalBusiness JSON-LD carrying legal name, address, phone, email, service area, founding date, and sameAs links to every profile you control.
• Service, Article, FAQPage, and BreadcrumbList schema on the pages where each genuinely applies — and nowhere else, because invalid markup is worse than none.
• Identical name, address, and phone across the site, the schema, your business profile, directories, and social pages. Consistency is what turns scattered mentions into one recognized entity.
• llms.txt at the domain root: a plain-language statement of who you are, what you do, and which pages are authoritative.
• A citable facts file — the numbers and attributes you want quoted correctly, published as JSON instead of buried inside a hero image.
• An entity graph and a JSON feed of new content, so a machine can enumerate your organization and its updates without scraping page HTML.

Layer 3 — write answers, not copy

The last rule is an architecture decision, not a style guide. A governed question library — each answer stored once and rendered everywhere it belongs — is how a site with hundreds of pages stays internally consistent. The alternative is four pages describing the same service four slightly different ways, which is exactly the ambiguity that costs citations.

• One page, one question, phrased in the heading the way a buyer would actually ask it.
• Answer within the first 40 to 60 words, before context and qualification. Retrieval favors the top of the page.
• Write self-contained sentences. Name the subject instead of writing we or this, so a quoted fragment still identifies your company.
• Prefer specifics to adjectives. Platforms, standards, numbers, constraints, and honest limits are quotable; superlatives are not.
• State the poor fits and the limits openly. Text that reads as evidence rather than promotion is treated better by models and by buyers.
• Keep one canonical answer per question across the whole site, reused wherever the question appears, so the site cannot contradict itself as it grows.

Layer 4 — measure honestly

Rank tracking does not describe generative visibility. Prompt testing does.

1. Write 30 to 60 buyer questions covering your services, your geography, and the comparisons you tend to lose.
2. Run them on a fixed schedule across the assistants your buyers actually use, from clean sessions without personalization.
3. Record four results per prompt: mentioned or not, cited with a link or not, described accurately or not, and which competitors appear beside you.
4. Check server logs for GPTBot, ClaudeBot, PerplexityBot, and Google-Extended fetches. A crawler that never visits explains every other symptom.
5. Track referral sessions from AI sources in analytics, accepting that attribution is incomplete — many buyers read an answer and type your name into a browser hours later.

A realistic first 90 days

The order is not negotiable. Answer content published on a site that AI crawlers cannot read is invisible work, and a flawless entity layer on a site with nothing worth quoting earns nothing.

1. Weeks 1-2, audit: fetch key pages as an AI crawler, inventory and validate structured data, check fact consistency across the site and third-party profiles, and run a baseline prompt panel so change can be proven later.
2. Weeks 3-6, rendering: prerender or server-render every public route, then verify content, meta tags, and JSON-LD in the raw HTML rather than in the browser inspector.
3. Weeks 5-8, entity layer: Organization and LocalBusiness schema, name and address cleanup everywhere, llms.txt, a citable facts file, an entity graph, and a content feed.
4. Weeks 7-12, answer layer: ten to twenty question-shaped pages answering what buyers actually ask, governed by one canonical answer each.
5. Week 12 onward: re-run the prompt panel monthly and treat every gap as a content backlog item rather than a mystery.

Mistakes that cost real money

• Buying guaranteed AI rankings or citations. No vendor controls another company's model or retrieval stack, so a guarantee is a sales tactic rather than a capability.
• Publishing bulk AI-generated pages. Volume without originality is deduplicated before it is ever quoted, and it dilutes the pages that would have earned citations.
• Serving special content to crawlers. Cloaking risks the classic rankings you already have, for no gain in generative visibility.
• Treating llms.txt as the project. It is a short file; rendering, schema, and consistency are the actual work.
• Leaving the CMS or API rendering path unfixed, so every new post ships invisible while the older static pages test fine.
• Assigning the whole thing to marketing. Rendering, routing, schema, and build validation are engineering decisions and will not fix themselves.

Where Geek Engineering fits

We treat this as software engineering, because the first two layers are exactly that: rendering pipelines, structured data generation, build-time validation, and content architecture. It is the same discipline we apply to industrial data — capture it correctly, give it context, make it trustworthy — pointed at the data a company publishes about itself.

This site is the working example. Every public route is prerendered to static HTML at build time; pages ship Organization, WebSite, WebPage, and Breadcrumb schema with Service, Article, and FAQPage markup where each applies; llms.txt, ai-info.txt, facts.json, knowledge-graph.json, ai-questions.json, and a JSON content feed are published at this domain root; and the build fails if the page registry, the internal link graph, or the structured data breaks. Every one of those claims is checkable from your browser without asking us for a report.

A typical engagement starts with an audit and a baseline prompt panel, moves to rendering and entity fixes, then builds the answer layer against the questions your buyers actually ask. What we will not sell is a promised position inside someone else's model — that is not something anyone can deliver, and vendors who promise it are describing luck.

SEO, AEO and GEO compared