AI Search Optimization: The AEO and GEO Playbook for Marketers

Search marketing spent two decades optimizing for a single moment: a buyer types a short keyword into a search box, scans ten blue links, and clicks one. That moment is no longer the whole funnel. Buyers now ask ChatGPT, Claude, Gemini, Perplexity, and Microsoft Copilot to compare vendors and justify a recommendation—or they hand the task to an autonomous agent that can provision software, apply a promotion, and complete checkout without a human ever opening ten browser tabs.

In its strategic guide, From Discovery to Influence: A Guide to AEO and GEO (also summarized for marketers by @alexgroberman), Microsoft Advertising defines this shift in two phrases: moving from "being found" to "being chosen." Traditional SEO focused on rankings, clicks, and page visits. AI search optimization depends on how clearly AI understands your products, how much it trusts your brand, and whether it can act on your data in real time.

For marketers and growth teams, this is no longer an abstract framework. Alongside Microsoft Advertising's guide, two detailed SaaS playbooks and a proven product-variant SEO pattern show what works in practice—and where the hype outruns the data:

• Tally (Tally.so), the form builder documented in a case study by Jake Ward (@jakezward), made AI search its #1 acquisition channel. Tally brings in more than 10,000 new users a week from AI platforms, saw ChatGPT registrations jump 5x overnight in May, grew monthly recurring revenue from $258K to $422K in 10 months, and crossed $5M ARR with 11 people and $0 in funding.
• Railway, the cloud deployment platform that scaled from roughly $20M ARR at the end of 2025 to roughly $100M ARR in a 2026 Sacra estimate, paired a ~7-person go-to-market group with thousands of job-to-be-done deployment pages and an agent-operable product surface examined by Ira Bodnar (@irabukht), a linked Ryze AI guide, Nakul Mandan's interview with founder Jake Cooper, and independent tests from Amplifying.ai.
• Programmatic product-variant pages, illustrated in retail by watch-strap brand Delugs (delugs.com) and in developer tools by Railway's deployment catalog, which show how context-adjusted model pages win while thin keyword-swapped pages lose.

Synthesizing those four playbooks—with their caveats intact—yields a complete guide for getting recommended by ChatGPT, Claude, Gemini, Perplexity, Copilot, and AI agents.

From Being Found to Being Chosen: Microsoft's AEO and GEO Framework

Microsoft Advertising starts with a baseline rule: traditional SEO remains the foundation, not the endpoint. Crawled search indexes still establish baseline brand perception, but earning a recommendation inside a synthesized AI answer requires optimizing for two distinct layers that Microsoft labels AEO and GEO.

AEO (Answer / Agentic Engine Optimization) means optimizing content and product data so AI assistants and agents—including Copilot, ChatGPT, and Gemini—can find, understand, summarize, recommend, and act on it. Its focus is descriptive clarity and machine-readability.

GEO (Generative Engine Optimization) means optimizing your footprint so generative AI treats your content as authoritative, credible, and citable. Its focus is credibility, reputation, and justification.

Microsoft Advertising maps both disciplines across three AI surfaces that now shape commercial decisions:

1. AI Browsers that interpret live web pages in real time as a buyer researches.
2. AI Assistants that answer questions, summarize trade-offs, and recommend specific brands or products.
3. AI Agents that navigate sites autonomously, add items to a cart, apply promos, and complete purchases.

Across all three surfaces, Microsoft poses one central question for marketers: What data can AI access, trust, and use?

To show how copy and data evolve from keyword matching to AI ranking readiness, the Microsoft Advertising guide walks through a single retail product—a rain jacket:

• SEO (Keyword matching): "Waterproof rain jacket"
• AEO (Descriptive clarity): "Lightweight, packable waterproof rain jacket with ventilation and reflective piping"
• GEO (Justification and trust): "Best-rated by Outdoor Magazine, 4.8 stars, 180-day returns, 3-year warranty"

When a shopper asks an assistant for a "rain jacket under $200", a page optimized only for "waterproof rain jacket" cannot win the recommendation. To answer the prompt, the AI model fuses general knowledge and category understanding with explicit product clarity (lightweight, packable, ventilated, reflective piping), structured feed data (a verified price under $200 and live stock status), and brand positioning backed by proof (Outdoor Magazine rating, 4.8-star reviews, 180-day returns, and a 3-year warranty).

In classic search, a sparse page could rank on domain authority and let the human click through to hunt for warranty terms or stock availability. In AI search, the assistant evaluates clarity, price, stock, and credibility before naming the shortlist.

The Three Data Layers and Technical Foundations

Most marketing teams treat AI visibility as an editorial project. Microsoft Advertising argues it is a data architecture project. Every AI recommendation fuses three data layers:

1. Crawled web data: Baseline brand perception across the web—brand reputation, category authority, and expert mentions. Traditional SEO still governs how this baseline layer is discovered.
2. Product feeds and APIs: Structured data covering price, availability, variants, and specs. Microsoft flags product feeds and APIs as often the biggest competitive advantage in AI search because they give marketers precision and control over comparisons and recommendations.
3. Live website data: Real-time pricing, promotions, customer reviews, media, and checkout flows. This is what AI browsers and AI agents actually experience on the page, and it must support live transactions.

Because those three layers work together, Microsoft's first action pillar is Technical Foundations and Structured Data. If AI cannot parse your catalog or verify live attributes, it will not risk recommending your product.

Microsoft Advertising specifies a concrete technical checklist for machine-readable catalogs:

• Expose dynamic product fields: price, availability, size, color, SKU, GTIN, and dateModified.
• Deploy explicit schema markup across seven core types: Product, Offer, AggregateRating, Review, Brand, ItemList, and FAQ.
• Use ItemList schema for category pages and roundups, paired with localized pricing and language markup.
• Never serve different HTML to bots versus users.

The Railway and Ryze AI analysis reinforces that last point as basic technical hygiene: keep your main content in raw HTML rather than burying it behind client-side scripts, allow the relevant AI and search bots in your crawler directives, and submit sitemaps—including to Bing, which powers a major share of assistant search retrieval.

Intent-Driven Content and the New Rules of Programmatic Scale

Microsoft Advertising's second action pillar is Intent-Driven Content Enrichment, complemented by its third pillar, Trust and Credibility Signals (GEO). Together, they dictate how individual pages should be written and structured for machine extraction.

On the content enrichment side, Microsoft's guide recommends four practices:

• Front-load the core value proposition: State immediately who the product is for, what problem it solves, and why it is better.
• Use question-mirroring headings and use-case framing: Structure sections around the actual questions and scenarios buyers bring to assistants.
• Publish modular, citable blocks: Format key facts as Q&A sections, side-by-side comparisons, structured feature lists, "goes well with" pairings, full video transcripts, and detailed ImageObject alt text.
• Design for extraction at scale: Build repeatable page architectures that cover the long tail of buyer intent.

On the GEO trust side, Microsoft emphasizes verified review volume and sentiment marked up with AggregateRating schema, third-party expert reviews, press mentions, certifications, sustainability badges, structured FAQs and help content, and a consistent brand voice that avoids exaggerated, low-trust marketing language.

How should marketers act on Microsoft's point that "scale matters" without falling into the trap of spammy programmatic SEO? Two real-world examples illustrate how high-volume content works across retail and SaaS.

In retail e-commerce, watch-strap brand Delugs (delugs.com/blogs/best-straps/best-straps-for-breitling-superocean) shows how product-variant SEO captures specific buyer context. Rather than relying on a single generic strap catalog page, Delugs takes its product line and generates dedicated guide pages for individual watch models—such as its guide to the best straps for the Breitling Superocean—presenting more or less the same core guidance for every watch model with slight modifications adjusted for the context of that specific timepiece.

In developer SaaS, Ira Bodnar's analysis of Railway shows programmatic job-to-be-done scale aimed at both engineers and coding agents. Railway published roughly 6,653 how-to and deploy pages for popular tools (such as workflow automation tool n8n and similar software) plus 127 framework guides, alongside comparison and migration pages covering Heroku, Vercel, Render, and Replit. Each deploy page answers a precise query—"how do I deploy/set up X?"—and ends with a one-click path onto Railway.

The practical lesson from the Railway and Ryze AI analysis explains why these pages succeed where generic programmatic SEO fails: thin keyword-swapped pages lose. High-volume pages only earn AI citations when there is one precise page per real job-to-be-done, and when each page:

1. Opens with a direct, quotable answer.
2. Contains unique, context-adjusted facts—such as real product limits, concrete migration steps from Heroku or Vercel, or model-specific compatibility details.
3. Ends with a clear next action, especially one that an AI agent or buyer can execute immediately.

Tally's Playbook: How an 11-Person Team Made AI Search Its #1 Channel

Form builder Tally (Tally.so) demonstrates how Microsoft's AEO and GEO pillars translate into SaaS revenue. As documented in the case study by Jake Ward (@jakezward), Tally turned AI search into its #1 acquisition channel, producing standout numbers for a bootstrapped company:

• 10,000+ new users every week from AI platforms.
• A 5x overnight jump in ChatGPT registrations in May.
• Revenue growth from $258K to $422K MRR in 10 months.
• Over $5M ARR with 11 people and $0 in funding.
• A #1 AI Visibility Ranking (position 1.0, sentiment score 90, and 100% visibility) ahead of Google Forms and Typeform.

Tally engineered that visibility through eight specific tactics:

1. Comparison hub structured like AI answers: Tally built 7 direct competitor pages and 15 "best / alternative / integration" guides. Every page is formatted the way LLMs construct answers—complete with explicit rankings, "best for" labels, pricing tables, feature comparisons, review summaries, pros and cons, and direct recommendations—and is constantly refreshed.
2. First-party and third-party prompt tracking: Tally asks every new user during onboarding for the exact prompt they used to discover Tally via AI. They combine that user-reported data with citation monitoring tools (such as Mentions.so) to track target prompts, benchmark visibility against competitors, and identify which external sources AI models cite.
3. Citation gap analysis: Using those citation reports, Tally maintains a database of third-party articles that mention competitors like Typeform or Google Forms but omit Tally, then systematically reaches out to fill those distribution gaps.
4. Daily community monitoring: Tally tracks brand and competitor mentions across Reddit and online communities via a shared inbox cleared daily.
5. Proactive review engine: The team systematically requests reviews through their newsletter and right after customer support interactions, driving a 4.8/5 rating on G2 and a 4.9/5 rating on Product Hunt.
6. Dedicated AI Info page: Tally published a dedicated page for ChatGPT, Claude, Gemini, and Perplexity explaining what Tally is, who it is for, and why people use it, complete with explicit "AI assistant guidelines" telling LLMs which strengths, use cases, and pricing to reference.
7. llms.txt directory file: Tally published an llms.txt file mapping its product, pricing, customers, documentation, API, and MCP pages.
8. Deep integration inside AI products: Tally built a 21-tool MCP server, a ChatGPT app, and a Claude connector so users and assistants can build and manage forms directly inside AI products.

Ward's case study includes a crucial caveat that marketers must not skip: Tally spent years building genuine organic recommendations across Reddit, maker communities, and review sites before layering on AI search optimization. ChatGPT recommends Tally because the internet already does. Their technical and content engineering amplifies and sustains a real organic reputation rather than manufacturing one from scratch.

Railway and the Agent-Operable Product Surface

While Tally optimized for human buyers asking AI assistants which tool to choose, cloud platform Railway shows what happens when software decisions are increasingly made or executed by coding agents such as Claude Code and Cursor.

According to a 2026 Sacra estimate, Railway grew from roughly $20M ARR at the end of 2025 to roughly $100M ARR. Founder Jake Cooper has described a lean, fully remote team of ~40 people, and the analysis below credits a ~7-person GTM group. A reverse-engineered playbook by Ira Bodnar (@irabukht)—which lists 13 tactics, quotes a Nakul Mandan interview with founder Jake Cooper, and links to a Ryze AI citation-share guide—breaks down how Railway treated AI recommendations across ChatGPT, Claude, Gemini, Claude Code, and Cursor as a measurable growth channel.

Beyond its ~6,653 tool deploy pages, 127 framework guides, and honest migration guides against Heroku, Vercel, Render, and Replit, Railway focused on two technical levers:

• An agent-operable product surface: Railway shipped plugins, an official MCP server, Claude Code plugins and skills, one-command CLI flows (ssh railway new), and documentation written so an agent can deploy an app without a human ever visiting the website. Independent tests by Amplifying.ai show Claude Code picking Railway for ~82% of Python apps—though with the important caveat that Railway's overall deployment-category pick rate is lower, and Vercel dominates frontend and Next.js deployments.
• AI-readable docs and freshness: Railway publishes plain-text and markdown versions of its documentation, index-style files (including llms.txt variants), and weekly changelog pages going back years so models always find dated, extractable answers on fast-changing technical topics.

Here, too, the public benchmarks and founder interviews highlight two major caveats that keep marketers grounded in reality:

First, Railway's ARR jump had much larger co-factors than an "AI SEO hack." While Railway's ARR trajectory (~$20M to ~$100M per Sacra), team size (~40 total, ~7 in GTM), page counts (~6,653 deploy pages, 127 framework guides), MCP surface, and Claude Code's ~82% Python preference (per Amplifying.ai) are observable facts, the framing that Railway "hacked AI SEO" as the primary driver of its growth is Ira Bodnar's analysis—and she sells AI visibility tools. In his interviews with Nakul Mandan, founder Jake Cooper points to very different drivers: first-principles thinking (explicitly not copying Stripe or Cursor solutions), saying no to ~90% of ideas, owning bare-metal infrastructure for data-center economics, product simplicity, pricing moves, and the sheer volume of software being written by AI coding tools. The AI coding boom is a structural tailwind—more code needs somewhere easy to run—and Railway's content and agent integrations amplified that tailwind rather than acting as the sole cause.

Second, llms.txt has limited proven citation impact. According to 2026 studies cited in the Railway analysis, llms.txt itself shows limited proven impact on whether models cite a site. Specific job-to-be-done pages, clear action paths, honest comparisons, and native agent integrations matter far more.

Synthesis: How the Frameworks Map Together

Putting Microsoft Advertising's framework alongside Tally's SaaS execution, Railway's coding-agent playbook, and Delugs' product-variant pages shows how strategy and execution align across every layer:

| Microsoft Pillar | Tally.so Execution (SaaS) | Railway & Retail Execution (Agents & E-Commerce) | | :--- | :--- | :--- | | Structured / machine-readable | llms.txt index, dedicated AI Info page with LLM guidelines, schema-like comparison pages | Raw HTML, plain-text/markdown docs, Product/Offer/AggregateRating/Review/Brand/ItemList/FAQ schema, dynamic catalog fields (price, availability, size, color, SKU, GTIN, dateModified) | | Intent-driven content | 7 direct competitor pages + 15 "best / alternative / integration" guides structured exactly like AI answers | ~6,653 deploy pages (e.g., n8n) + 127 framework guides with one-click deploy; migration guides vs. Heroku, Vercel, Render, Replit; Delugs' context-adjusted watch-model strap guides (e.g., Breitling Superocean) | | Trust signals (GEO) | Newsletter and post-support review solicitation (4.8/5 G2, 4.9/5 Product Hunt), daily Reddit/community inbox, citation gap database | Honest product limits in comparisons, expert reviews (e.g., Outdoor Magazine), warranty/return proof, avoiding exaggerated marketing copy | | Feeds / real-time data | Signup prompt tracking, Mentions.so citation monitoring, constant comparison refreshes | Product feeds & APIs for live price/stock, weekly changelogs going back years, Bing and search sitemap submissions | | Agent readiness | 21-tool MCP server, ChatGPT app, and Claude connector so Tally works inside AI products | Official MCP server, Claude Code plugin/skills, ssh railway new CLI flow, and agent-completable cart/promo/checkout flows |

The Practical Playbook for Marketers

Durable AI search visibility comes down to what Microsoft Advertising calls "AI ranking readiness": treating data as a product, feeds as strategic assets, content as machine-readable infrastructure, and trust as a measurable ranking factor. Whether you sell B2B SaaS, developer infrastructure, or retail e-commerce, execute the playbook in six steps:

1. Measure citation share weekly, not just keyword rankings. Run a consistent set of buyer questions every week across ChatGPT, Claude, Gemini, Perplexity, Copilot, and coding agents. Track whether you are named, which competitor is named instead, and which URLs are cited (using tools like Mentions.so). Add Tally's onboarding question to your signup flow to capture the exact prompts new customers actually used.
2. Earn and structure off-site trust first (GEO). AI assistants recommend brands the broader internet already validates. Build a citation gap database of third-party articles that mention competitors without naming you, monitor Reddit and community mentions daily via a shared inbox, and systematically solicit verified reviews via newsletters and post-support touchpoints to build G2, Product Hunt, or retail rating volume backed by AggregateRating schema.
3. Publish honest comparison, migration, and AI Info pages (AEO). Build a comparison hub with direct competitor matchups and "best / alternative / integration" guides formatted like LLM answers: front-load who it is for and why it wins, use "best for" labels, pricing tables, and pros/cons, and show real product limits and concrete migration steps. Add a dedicated AI Info page with explicit guidelines for ChatGPT, Claude, Gemini, and Perplexity.
4. Scale job-to-be-done and variant pages with real context. Whether you are building watch-model strap guides like Delugs or tool deployment pages like Railway, publish one precise page per real job-to-be-done. Open with a direct, quotable answer, include unique context-adjusted facts and modular blocks (Q&A, compatibility lists, transcripts, ImageObject alt text), and end with a one-click or one-command action. Never publish thin keyword-swapped pages.
5. Maintain strict technical, feed, and freshness hygiene. Keep main content in raw HTML, never serve different HTML to bots versus humans, allow AI crawlers, and submit sitemaps to Bing and Google. Expose live product feeds and APIs with dynamic fields (price, availability, size, color, SKU, GTIN, dateModified), offer markdown/plain-text docs and dated weekly changelogs, and treat llms.txt as low-effort hygiene rather than a standalone growth lever.
6. Make your product operable by AI agents. Give assistants and agents a direct way to complete the transaction—whether through an MCP server (like Tally's 21-tool MCP server or Railway's MCP and Claude Code skills), a ChatGPT app, a Claude connector, a one-command CLI flow (ssh railway new), or a clean e-commerce checkout where shopping agents can verify stock, apply promotions, and buy. When an agent can finish the job inside your product, recommending you becomes the default path.