Generative Engine Optimization (GEO): How to Get Cited by AI in 2026

Generative engine optimization (GEO) is the practice of making your content easy for AI answer engines (ChatGPT, Google AI Overviews and AI Mode, Perplexity, Claude, Copilot) to find, trust, and cite when they write an answer. Classic SEO wins you a ranking; GEO wins you a mention or a citation inside a synthesized answer that often replaces the ten blue links. It is not a separate discipline with secret tricks: it is SEO done for a reader that is now a language model with a search tool, plus a handful of things that specifically raise your odds of being quoted, and a few popular "hacks" that measurably do nothing.

Key takeaways

• GEO = being retrievable + being quotable. AI engines run a search first, then pick a few passages to ground the answer. You need to show up in that search and then give the model something concrete to cite: a clear claim, a number, a definition, a comparison.
• The traffic math has changed. Ahrefs measured a 34.5% drop in clicks to top-ranking pages when an AI Overview appears (and up to 58% in its larger late-2025 follow-up), and Pew found only 1% of users click a source cited in an AI summary. Visibility in the answer itself is now part of the prize.
• Google says GEO is "still SEO." Its May 2026 guide for generative AI features in Search is explicit: no special markup, no chunking, no llms.txt needed. Unique, non-commodity content and a clean technical setup are what matter.
• For ecommerce, the feed is the page. ChatGPT shopping and Google AI Mode pull products from merchant feeds (Shopify's catalog, Merchant Center), so price, stock, shipping and attribute accuracy drive visibility more than a clever product description.
• Measure before you optimize. Track AI referrals in GA4, run a fixed prompt set monthly, and only then decide whether a paid AI-visibility tool is worth it.

What generative engine optimization actually is

When someone asks ChatGPT "what's the best running shoe for flat feet under $150," the model doesn't answer from memory alone. For most queries it fires off one or several web searches, pulls back a list of candidate pages, extracts the relevant passages, and writes an answer that cites a handful of them. Google AI Overviews and AI Mode do the same thing on top of Google's index; Perplexity does it on its own index plus live search.

That pipeline has three gates, and GEO is about passing all three:

• Retrieval: your page has to come back in the engine's internal search. This is almost entirely classic SEO: indexable, relevant, authoritative enough to rank for the query the model generated (which is often a rephrased, longer version of what the user typed).
• Selection: out of 10 to 30 candidates, the model picks the passages that answer the question most directly. Clear, self-contained paragraphs with specific facts win here.
• Citation: the engine attaches your URL as a source, or names your brand in the text. Both count: a brand mention without a link still shapes what the user buys next.

GEO vs SEO vs AEO

You'll see three acronyms thrown around. Here is the honest version:

• SEO gets you ranked in a list of results. Still the foundation, because retrieval runs on the same index.
• AEO (answer engine optimization) is the older term for winning featured snippets and voice answers: one short, direct answer to one question.
• GEO (generative engine optimization) targets synthesized, multi-source answers. The difference from AEO is that you are competing to be *one of several* sources, so comparisons, data and unique perspective matter more than a single tidy sentence.

Google's own documentation, published in May 2026, calls AEO and GEO terms "you may see used" and states that optimizing for generative AI search is still SEO. Treat the label as marketing; treat the mechanics as real.

Why generative engine optimization matters in 2026

The numbers moved fast:

• At Google I/O in May 2026, Google said AI Mode passed 1 billion monthly users and AI Overviews reached 2.5 billion, and the two surfaces were merged into one search flow.
• Ahrefs' April 2025 study found AI Overviews cut clicks to the top-ranking result by 34.5% for informational queries; its December 2025 update on 300,000 keywords found the loss reaching 58%.
• Pew Research's analysis of 68,879 Google searches found that just 1% of users who saw an AI summary clicked a cited source, versus 15% who clicked a traditional result when no summary appeared.
• For shopping specifically, BrightEdge's holiday 2025 analysis found only about 4% of AI Overview shopping citations went to retailers, compared with roughly 36% in ChatGPT. Different engines, very different odds for a store.

The practical conclusion: you can't fully replace lost clicks, but you can decide whether your brand is *in* the answer or absent from it. For a store, being the product the assistant recommends is worth more than the click used to be.

How generative engines choose what to cite

The original GEO research paper (Princeton and collaborators, published at KDD 2024) tested dozens of content tweaks on a benchmark of real queries. The ones that lifted visibility in generated answers were adding statistics, adding quotations from credible sources, and citing sources, each worth a double-digit percentage gain in share of the answer. Keyword stuffing did nothing or hurt. Two years later, the pattern still holds across independent analyses, with a few additions:

• Answer-first structure. The first 1–2 sentences under a heading should answer the heading. Models extract passages, not pages.
• Specificity over fluency. "Our tests showed a 12% faster drain at 40 PSI" is citable; "our hose drains fast" is not.
• Third-party corroboration. Engines cross-check. If review sites, Reddit threads, Wikipedia or trade publications mention you in the same context, you become a safer source to cite. This is why PR, reviews and community presence now show up in GEO plans.
• Freshness. Dated content with a visible "updated" date and current figures gets picked over stale pages for time-sensitive queries.
• Being crawlable by the right bots. If OAI-SearchBot or PerplexityBot is blocked, you're out of that engine's retrieval entirely.

What does *not* work, per Google's 2026 guide and the largest studies: llms.txt files (SE Ranking's analysis of 300,000 domains found no citation difference, and Google states it ignores them), "chunking" content into tiny fragments, rewriting your pages "for AI," special schema types, and buying inauthentic mentions.

Step 1: Audit your current AI visibility

Before changing anything, find out how engines already talk about you. Do it by hand first, so you understand what the engines see.

By hand

• Build a prompt set. Write 20–30 questions a real customer would ask in each stage: "what is X," "best X for Y," "X vs Z," "is [your brand] legit," "[your brand] alternatives." Keep them in a spreadsheet.
• Run each one in four engines: ChatGPT (with search on), Google AI Mode, Perplexity and Claude. Use a logged-out or fresh session where possible so personalization doesn't skew results.
• Record three things per answer: was your brand mentioned, was your URL cited, and which competitors were cited. Note the exact page cited (it's often not your homepage).
• Open the cited competitor pages and look at what the engine quoted: usually a stats line, a comparison table or a clearly worded definition near a heading. That's your template.
• Check GA4 for AI referrals: Reports → Acquisition → Traffic acquisition, then filter *Session source* for chatgpt.com, perplexity.ai, copilot.microsoft.com, gemini.google.com and claude.ai. Low numbers are normal; the trend is what you track.

Faster with AI

Paste your prompt set into Claude or ChatGPT and ask it to produce the spreadsheet and the first analysis for you. A prompt that works:

> You are an AI-search analyst. Here is my brand ([brand]), my category ([category]) and 25 customer questions (list). For each question, write the exact query a shopper would type into ChatGPT, Google AI Mode and Perplexity. Then give me a scoring sheet with columns: question, engine, brand mentioned (Y/N), URL cited (Y/N), competitors cited, and "what the cited passage looked like." Output as a table I can paste into Google Sheets.

With Claude Code you can go further: give it a CSV of your prompt set and the pasted answers, and ask it to cluster the questions where you're absent by topic and by content type ("comparison," "how-to," "spec sheet"), so you see which gaps are fixable with one page and which need a series.

Step 2: Make your pages citable

This is the core of GEO, and it's mostly editing, not rebuilding.

By hand: the citability checklist

Open the page you want cited for a specific question and go through this list:

• Lead with the answer. Under every H2, put a 40–60 word paragraph that fully answers the heading's question. If someone read only that paragraph, they'd have the gist.
• Add two or three numbers you can stand behind. Test results, prices, dimensions, time saved, percentages from your own data. Cite where each number comes from.
• Add one quote from a credible source. A named expert, a standards body, a published study. Link it.
• Add a comparison. A simple table ("Model A vs Model B: weight, battery, price, best for") is one of the most frequently quoted formats in shopping answers.
• Use the words people use. Match the question's phrasing in the heading ("Is merino wool itchy?" rather than "Comfort properties").
• Show the author and the date. A real name, a role, an "updated" date. Engines and users both read them.
• Cut everything that's generic. If a paragraph could appear on any competitor's site unchanged, it's commodity content and it isn't helping you.

A concrete example: a page titled "Best standing desks for small apartments" that says "Our favorite is the X, it's great for small spaces" will lose to one that says "The X fits a 48-inch wall, weighs 62 lb, lifts in 11 seconds in our test, and costs $389 as of October 2026. Compared with the Y (52-inch, $449), it gives up 4 inches of width for $60 in savings." The second is quotable as-is.

With Claude or ChatGPT: a rewrite that keeps your voice

Don't ask the model to "rewrite for AI." Ask it to find the gaps and draft the specific pieces you then verify. Prompt:

> Here is a page from my site (paste text) and the question I want it cited for: "[question]". Audit it against this checklist: answer-first paragraph under each H2, at least 3 specific statistics with sources, 1 expert quotation, 1 comparison table, headings phrased as the user's question, visible author and date. For every gap, draft the missing element and mark clearly where I need to insert real data (use [NEEDS DATA]). Do not invent statistics. Keep my tone.

Then you fill in the [NEEDS DATA] markers with real figures. The model's job is structure and drafting; the facts are yours. If you have a catalog of 500 pages, Claude Code can run this audit across a folder of exported HTML and output a prioritized list: pages with the biggest gaps that also have the most impressions in Search Console.

Step 3: Technical GEO (what to check once)

Most of this is standard SEO hygiene, with a few AI-specific twists.

• Let the right crawlers in. Check your robots.txt for OAI-SearchBot (ChatGPT search), ChatGPT-User (on-demand fetches when a user asks), PerplexityBot, ClaudeBot and Googlebot. Many sites blocked "all AI bots" in 2023–2024 to avoid training and accidentally blocked search too. GPTBot is OpenAI's *training* crawler; OAI-SearchBot is the one that matters for citations. Google-Extended controls Gemini training and does not affect AI Overviews or AI Mode, which use Googlebot.
• Render content server-side. Some AI crawlers execute little or no JavaScript. If your key copy, specs or prices only appear after a client-side render, they may be invisible to the engine.
• Keep structured data, don't obsess over it. Product, Article, FAQ and Organization schema help engines parse price, availability and authorship. Google's guide says structured data isn't required for generative features and there's no special schema to add, so treat it as a parsing aid, not a lever.
• Fix duplicates and canonicals. Models that find five near-identical versions of a page tend to cite none of them, or the wrong one.
• Speed and stability. Timeouts during an on-demand fetch (ChatGPT-User, Perplexity) mean the engine simply uses another source.
• Skip llms.txt unless you ship developer docs. It's harmless, but the evidence says it doesn't move citations, and Google ignores it.

Step 4: Generative engine optimization for ecommerce

For a store, product discovery in AI engines increasingly comes from feeds, not from crawled pages.

• ChatGPT: OpenAI launched Instant Checkout in September 2025 with the Agentic Commerce Protocol (built with Stripe), then in March 2026 shifted toward discovery and comparison with checkout on the merchant's own site. Shopify merchants are syndicated through Shopify's catalog; other stores can submit a product feed directly. What the assistant shows is driven by title, attributes, price, availability, shipping and returns data.
• Google AI Mode / AI Overviews: product results come from the Shopping Graph, which is fed by Merchant Center. Google and Shopify launched the Universal Commerce Protocol in January 2026 for agentic checkout. Google's optimization guide explicitly tells ecommerce sites to keep Merchant Center data accurate as part of generative-AI readiness.
• What to fix first: feed completeness (GTIN, brand, color, size, material), price parity between feed and page, in-stock accuracy, and shipping/return policies that are machine-readable. A product with a complete feed and a boring description will beat a beautifully written page with a stale price.
• Then content: buying guides and comparisons on your own domain are what engines cite for "best X for Y" questions, which is where new customers enter. Product pages themselves are rarely the cited source in Google's AI surfaces, so the guide page is your GEO asset and the feed is your shopping asset.

A manual check: ask ChatGPT and AI Mode for "best [your category] under $[price]" and see whether your products appear in the carousel, with what title and price. If the price is wrong or the product is missing, the fix is in the feed, not the copy.

Step 5: Measure GEO without fooling yourself

• Monthly prompt set run (from Step 1): share of prompts where you're mentioned, cited, and your average "position" (first source vs fifth). This is your AI visibility score; keep the prompt set stable so it's comparable.
• GA4 referrals from AI domains, plus conversion rate of that traffic. It's small but typically converts well because the assistant already pre-qualified the user.
• Search Console for the pages you optimized: Google includes AI Mode and AI Overview impressions and clicks in the Performance report without a separate filter, so watch impressions holding up while clicks fall. That pattern means you're in the answer.
• Branded search volume in Google Trends or Search Console: brand mentions in AI answers without links tend to show up as more people typing your name.

AI visibility tools: real options, fairly compared

You can run everything above with a spreadsheet. If you need scale or client reporting, these are the main categories (pricing as of late 2026; verify with the vendor):

• Dedicated AI-visibility platforms (Profound, Peec AI, Otterly, Scrunch). Pros: automated prompt runs across many engines, competitor share-of-voice, citation tracking at scale. Cons: cost, and prompt sets that may not match what your customers actually ask. Profound's entry tier was listed around $99/month for a single engine and $399/month for three; Peec and Otterly start lower.
• SEO suites with AI add-ons (Semrush AI Toolkit, Ahrefs Brand Radar). Pros: lives next to your existing rank and keyword data. Cons: fewer engines covered, prompt-based estimates rather than your real customers' queries. Semrush's AI Visibility Toolkit was sold at about $99/month per domain.
• Do-it-yourself with Claude Code or a script. Pros: free beyond API costs, fully custom prompts. Cons: engines' consumer answers differ from API answers, so you're approximating.

Honest advice: start with the spreadsheet for three months. If you find yourself running the same 100 prompts by hand, that's when a tool pays off.

Mistakes that waste GEO budgets

• Publishing "AI-optimized" versions of pages. Google explicitly says you don't need to rewrite content for AI systems, and duplicate pages dilute you.
• Chasing every engine with the same tactic. ChatGPT cites retailers far more than Google's AI surfaces do. Prioritize by where your category actually gets cited.
• Buying mentions. Inauthentic mentions are called out in Google's guide and are easy for engines to discount.
• Measuring only clicks. If your impressions in Search Console held while clicks dropped, you're being read inside the answer. That's not a failure; it's the new baseline.
• Forgetting the feed. For stores, half of GEO is product data quality. No amount of blog content fixes a missing GTIN.

How savvy.so does it

savvy.so doesn't track AI citations, but it covers two pieces of the ecommerce side of GEO for Shopify stores. Its AI catalog tool writes product, collection and blog content starting from keyword research, with the attributes, SEO fields, photos and documents that both Google and AI assistants read, so your product data is complete and consistent across page and feed. It also sends your feed to Merchant Center and monitors competitor prices in the Google Shopping buy box, so the price an assistant shows is the one you intend. And because it ties Search Console pages to actual orders and profit, you can see which guide pages that AI engines cite actually make money, not just impressions. There's a free trial.

FAQ