Skip to content

GEO & AEO: a learning guide

Read this once, top to bottom (about 40 minutes). Then use the skill to apply it to one feature. Last checked: 18 Oct 2026. AI search changes fast: re-check the official pages in Sources every quarter.

Contents

  1. The words
  2. How an AI answer picks its sources
  3. What the official docs say
  4. What the research says
  5. The three pillars
  6. GEO before the product exists
  7. Crawler access
  8. Measuring it day by day
  9. Myths
  10. A 4-week study plan

1. The words

Term Plain meaning Where it shows up
SEO (search engine optimization) Get a page ranked in a list of links. Google, Bing classic results
AEO (answer engine optimization) Get a passage used as the direct answer to a question. Featured snippets, People Also Ask, voice assistants, AI Overviews
GEO (generative engine optimization) Get your page cited or your product named inside an answer an AI model writes. ChatGPT search, Perplexity, Google AI Overviews / AI Mode, Copilot, Gemini, Claude
LLMO / AI SEO Other names for GEO. Vendor blogs
Citation The answer links to your page as a source. All AI engines
Mention / recommendation The answer names your product, ideally on the shortlist. Mostly ChatGPT, Perplexity, Gemini
Grounding query The search the AI ran in the background to find sources. Bing AI Performance report
Query fan-out One question split into many background searches. Google AI Mode, ChatGPT

AEO and GEO overlap. In practice: AEO = write the answer so it can be lifted; GEO = be the source the model trusts enough to cite or recommend. Both sit on top of normal SEO. They do not replace it.

2. How an AI answer picks its sources

Most AI search products follow the same loop (retrieval-augmented generation):

User question
   │
   ▼
Fan-out: 3–20 background searches ("best X for Y", "X vs Z", "X pricing", "site:official")
   │
   ▼
Retrieve: pages from the engine's index (Google, Bing, own crawler)  ← you must be crawlable + indexed
   │
   ▼
Pick passages: short blocks that answer one sub-question clearly    ← AEO: structure
   │
   ▼
Write answer + cite the passages it leaned on                        ← GEO: authority, facts, consensus

Three things follow from this: - No index, no citation. If the engine can't crawl the page, nothing else matters. - Passages win, not pages. A clear 40–60 word block under a question heading is easy to lift. A wall of marketing copy is not. - Other sites count. Engines also retrieve Reddit threads, reviews, directories, YouTube text and news. What they say about you shapes whether you get recommended.

3. What the official docs say

These are the only "rules". Everything else in this guide is research or practice.

Google (AI Overviews, AI Mode). From AI features and your website (updated 10 Dec 2025): - To be shown, a page must be indexed and eligible to show a snippet. There are no extra requirements and no special optimization needed. - You don't need new machine-readable files, AI text files or special schema.org markup. - Clicks and impressions from AI features are counted in Search Console → Performance, search type "Web". There is no separate AI report. - Controls: nosnippet, data-nosnippet, max-snippet, noindex. Google-Extended controls use in Gemini training and grounding, not Search.

Google spam policies (Spam policies for Google web search). Three matter for feature previews: - Misleading functionality: pages that make users think they can get a service that they can't. A "Try it now" button for a feature that doesn't exist is exactly this. - Doorway abuse: many near-identical pages made to rank for similar queries. - Scaled content abuse: many pages generated mainly to manipulate rankings.

OpenAI (Overview of OpenAI crawlers): - OAI-SearchBot decides whether you can appear in ChatGPT search answers. Allow it in robots.txt and in your firewall. Changes take about 24 hours. - GPTBot is for model training. Blocking it does not remove you from ChatGPT search. - ChatGPT-User fetches a page when a user asks; it is not an automatic crawler.

Perplexity (Perplexity crawlers): allow PerplexityBot (search results, not training) and its published IPs; Perplexity-User fetches on a user's request.

Anthropic (About Claude's web crawlers): ClaudeBot (training), Claude-SearchBot (search quality), Claude-User (user-requested fetches).

Microsoft Bing (AI Performance in Bing Webmaster Tools, public preview, Feb 2026): the first first-party report of citations in Copilot and Bing AI answers (total citations, cited pages, grounding queries). It shows no clicks.

US FTC (Rule on consumer reviews and testimonials, 16 CFR Part 465, in force 21 Oct 2024): no fake or AI-written reviews, no undisclosed insider reviews, no "independent" review site you control. A pre-launch page has no reviews yet. Don't invent them.

4. What the research says

The GEO paper (Aggarwal et al., Princeton / IIT Delhi, KDD 2024, arXiv 2311.09735) tested 9 ways to rewrite a source page and measured how much of the AI answer used it. Visibility change vs. no change, on their GEO-bench (Table 1, word-count metric):

Rewrite Change
Add quotations from named people about +41%
Add statistics about +33%
Fluency (clearer, smoother writing) about +29%
Cite sources about +28%
Technical terms about +18%
Easy-to-understand language about +14%
Authoritative tone about +12%
Keyword stuffing about −9%

Two findings matter most for a small, new site: - Lower-ranked pages gain the most. For a page in 5th position, cite sources raised visibility by about +115%, while the page in 1st position lost share. GEO is one of the few levers where being small is not a handicap. - Results differ by topic. Test on your own queries; don't copy percentages into a forecast.

Treat the numbers as direction, not a promise: it was a lab setup on a research engine and Perplexity in 2023–24.

Industry practice (vendor studies, re-check before quoting): AI Overviews show on a large share of informational queries and lower click-through on the top link; ChatGPT's preferred page formats shifted during 2026 toward official and primary pages and away from listicles. The lesson that holds across studies: owned "official" pages with plain facts, plus real third-party mentions, beat thin listicles.

5. The three pillars: structure, authority, presence

Structure (make it liftable) — this is AEO

  • One page = one main question. H1 says what the thing is.
  • Under each H2 written as a question ("Does X work offline?"), put the answer first in 40–60 words, then detail.
  • Tables for comparisons, numbered lists for steps, one idea per paragraph.
  • Key text is in HTML, not inside images or behind a JavaScript-only render. Images get alt text.
  • A visible "Last updated" date and a named author.
  • Optional structured data that matches the visible text (SoftwareApplication, FAQPage, Article). Google says it's not required for AI features; it can help other engines and is normal good SEO.

Authority (make it worth citing) — this is GEO

  • Facts you own: numbers from your own tests, a dated design decision, a measured benchmark. "Exports a 4K frame in 1.8 s on an M2 Air in our test" beats "blazing fast".
  • Cite sources for every outside claim, with links.
  • Named people: quote the designer or a beta tester (with permission, and real).
  • Show your work: mockups, a design review log, a changelog. These are first-hand experience that a model can't get anywhere else.

Presence (be where the AI looks)

  • Answer real questions where people ask them (Reddit, forums, Discord, Stack Exchange) and link only when it helps.
  • Get listed in directories and "alternatives" pages that are already cited for your queries.
  • Short video with a written description and captions (the text layer is what gets read).
  • Say the same one-line positioning everywhere: About page, store listing, directories, profiles. Models look for agreement between sources.

6. GEO before the product exists

You can do GEO before writing code, because the thing an AI cites is a clear answer, not a working binary. What you publish is a feature preview: an honest page about something you are designing.

What a preview page can honestly offer - The problem, in the words people use when they search for it (from your research). - The design: mockups, a short clickable prototype video, the decision log from your design review. - What it will and won't do, and the status: In design · Not available yet · Planned for {{month}}. - A way to act: join the waitlist, vote, comment on the mockup, become a beta tester. - How it compares with what people use today (facts about competitors, cited).

What it must never do - Buttons like "Download" or "Start free trial" for something that doesn't exist (Google: misleading functionality). - Fake reviews, fake user counts, fake "as seen on" logos (FTC rule). - One page per keyword variant ("X for students", "X for teachers"… with the same text) (doorway / scaled content). - Screenshots presented as the real app when they are mockups. Label them "Design mockup".

The validation loop. Traffic tells you whether people are looking; sign-ups tell you whether they want it.

Research questions  →  1 preview page per feature  →  index + share in 3 places
        ▲                                                    │
        │                                                    ▼
 build / rework / drop  ◄──  read 7-day trend  ◄──  daily log: visits, AI visits, sign-ups

7. Crawler access (robots.txt)

Decide per purpose. A common setup for a new product that wants to be found but not used for training:

# Search and answer engines: allow
User-agent: Googlebot
Allow: /
User-agent: Bingbot
Allow: /
User-agent: OAI-SearchBot
Allow: /
User-agent: PerplexityBot
Allow: /
User-agent: Claude-SearchBot
Allow: /

# Model training: your choice (blocking these does not remove you from search answers)
User-agent: GPTBot
Disallow: /
User-agent: ClaudeBot
Disallow: /
User-agent: Google-Extended
Disallow: /

Sitemap: https://example.com/sitemap.xml

Then check that your host's firewall or bot protection (Cloudflare, Vercel) isn't blocking the allowed bots anyway.

llms.txt is a community proposal (llmstxt.org). Google says AI text files aren't needed. It costs 10 minutes and does no harm; don't expect traffic from it.

8. Measuring it day by day

What you can measure, and where

Signal Tool Delay What it tells you
Impressions, clicks, queries per page (incl. AI Overviews / AI Mode) Google Search Console → Performance → Web 1–3 days Google is showing the page, for which questions
AI citations in Copilot / Bing Bing Webmaster Tools → AI Performance days Bing's AI uses your page, with grounding queries
Visits from AI assistants GA4 (session source chatgpt.com, perplexity.ai, claude.ai, gemini.google.com, copilot.microsoft.com) same day People clicked through from an answer
Waitlist sign-ups, mockup votes GA4 key event + your form tool same day They want it
Mentioned or not Manual prompt checks (below) weekly The AI knows you exist

GA4 setup (10 minutes). Admin → Data display → Channel groups → create a custom channel "AI assistants" with Session source matches regex:

chatgpt\.com|chat\.openai\.com|perplexity\.ai|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com|bing\.com/chat

Put it above "Referral" (first match wins). ChatGPT usually adds utm_source=chatgpt.com to links; other assistants often send no referrer, so some AI visits will show as Direct. Your AI number is a floor, not the total.

Reading daily numbers without fooling yourself

  • Daily traffic on a new page is noise. 3 visits one day and 11 the next means nothing. Read the 7-day rolling average and compare it with the baseline (the 7 days before you published or changed the page).
  • One change at a time per page, and write the date in the log. Otherwise you won't know what worked.
  • Expect a lag. Indexing can take days to weeks; AI citations usually come later than Google impressions. Judge a page after 28 days, not 3.
  • Count rates, not single runs. AI answers vary. Run each prompt 3 times and record "mentioned 1/3".
  • Sign-ups per 100 visits is the number that decides whether to build. Traffic without sign-ups means the page answers a question but the feature isn't wanted (or the page doesn't ask clearly).

Weekly prompt check (15 minutes)

Pick 5 prompts a buyer would type (problem-aware: "how do I…"; solution-aware: "best app for…"; comparison: "X vs Y"). Run each 3 times in ChatGPT (search on), Perplexity and Google. Log: mentioned? cited page? who else is cited? The sites cited instead of you are your outreach list.

9. Myths

Myth Reality
"GEO replaces SEO." AI answers retrieve from search indexes. No index, no citation.
"You need llms.txt / special schema for Google AI." Google says no special files or markup are needed.
"Blocking GPTBot hides me from ChatGPT." Search inclusion is OAI-SearchBot. Training and search are separate tokens.
"Write separate pages for AI." Google calls mass AI-targeted variants scaled content abuse. Write for people, structure clearly.
"More keywords = more citations." Keyword stuffing lowered visibility in the GEO study.
"A preview page can't rank because the app isn't out." Engines cite clear answers and original material. Honest previews with mockups and decisions are original material.
"A spike after a Reddit post means GEO works." Separate channels in the log. GEO shows up as organic + AI sources over weeks.

10. A 4-week study plan

Week Learn Do (on your product)
1 Sections 1–3 Verify the site in Search Console and Bing Webmaster Tools; add the GA4 "AI assistants" channel; fix robots.txt; start the daily log.
2 Sections 4–6 Pick 1–3 features. Collect the real questions. Publish one preview page per feature with mockups and a waitlist.
3 Section 5 (presence) Answer 5 real threads; get listed in 2 directories cited for your queries; post a 30-second mockup walk-through with captions.
4 Sections 8–9 Run the weekly prompt check. Run traffic_trend.py on the log. Decide per feature: build, rework the page, or drop.

Sources