SEO, GEO, AEO: What Optimization Means When AI Answers First

I knew SEO. The terms GEO and AEO caught me by surprise; I had no idea they existed until a few months ago. And it's not like I study marketing: I've been writing code for years, but nobody in my dev bubble talked about this. When I started digging, I found two worlds: vendors renaming old services with a fresh invoice line, and serious studies with numbers that genuinely change how you think about a website. This post is the research distilled: what each term means, what the evidence says, and what's smoke.

The short promise: by the end you'll know which optimization applies to your site, what you can ignore guilt-free, and what to measure so you don't fool yourself.

You'll read 2-3 opinions about how much traffic Google is losing, all citing data. That's the first thing to sort out.

Clicks dropped and nobody agrees on how much

The hard numbers first:

• Pew Research tracked 68,879 real searches by 900 US adults: when an AI Overview appeared, people clicked a traditional result 8% of the time. Without one, 15%. Nearly half as much. - Source: Pew Research Center
• Ahrefs measured across 300,000 keywords: position 1 CTR fell from 7.3% to 1.6% when an AI Overview sits on top. A 58% drop. - Source: Ahrefs
• SparkToro, using Similarweb's panel: 68.01% of Google searches end with zero clicks (January-April 2026).

And Google? It denies the trend. Liz Reid, Google's head of Search, wrote in August 2025 that total organic click volume remains "relatively stable" year over year, that clicks are "higher quality," and that third-party studies use "flawed methodologies." - Source: Google blog. The odd part: the post doesn't include a single number.

Both things can be true at once: there are more searches than ever (Google says AI Mode passed one billion monthly users in May 2026), but each search produces fewer clicks. If your business depends on organic traffic, planning around "total volume is stable" means planning around the metric you care about least.

Three terms, three doors

Start with the genealogy, because it reveals how much is fashion and how much is substance:

• SEO (search engine optimization): the classic. Optimize to appear in Google's list of links and win the click.
• AEO (answer engine optimization): the term dates back to at least 2018, the featured-snippet and voice-assistant era. It means being the extracted answer, not just another link.
• GEO (generative engine optimization): coined in November 2023 by a Princeton paper - Source: Aggarwal et al., arXiv -, accepted at KDD 2024. The authors measured how editing a page changes its visibility inside LLM-generated answers.

The metaphor that ordered my head (credit to EMGI's analysis): an AI answer has two doors. The memory door: what the model learned during training about your brand. The retrieval door: what the model searches live on the web when it doesn't know or needs to verify. GEO works the memory door; AEO works the retrieval door. And SEO, increasingly, is what keeps the retrieval door open, because AI systems end up retrieving from pages that already rank.

In practice there's a third voice saying something else. Google's official guide for optimizing toward AI features, published in May 2026, defines both terms and then adds: "From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO". - Source: Google Search Central Translation: for Google this is SEO with another name, with no new technical requirements and no special schema. When a vendor sells you GEO as a separate discipline, keep in mind that the channel's owner says it isn't one.

So why study all three? Because the output differs: SEO competes for a position in a list and counts clicks. GEO and AEO compete to be one of the few sources the model cites, and count citations and mentions. The machinery underneath is the same: crawl → index → retrieve → rank, with an LLM at the end. What changes is the prize.

Ranking ≠ citation

Here's the number that surprised me most in the research: ranking and being cited are different games.

Ahrefs compared 15,000 AI prompts against Google rankings: only 12% of the URLs cited by LLMs also appear in Google's top 10. BrightEdge put it at 17% for AI Overviews and Moz at 12% for AI Mode. - Source: Ahrefs

EMGI's study (150 SaaS companies, 120 buying queries) put numbers on why:

• Brand authority predicts ChatGPT citations at r = 0.76. Organic traffic: r = 0.23. Nearly nothing.
• 44% of brands ranking in Google's top 10 are completely invisible to ChatGPT for the same queries.
• The extreme case: Customer.io tied HubSpot for the most ChatGPT citations (38 each) with roughly 125x less traffic.
• Source: EMGI

The explanation is in how models store knowledge. An LLM doesn't store keywords: it stores a semantic map of what appears together. Your brand is not what your website says it is; it's the consensus of what others write about you: editorial coverage, listicles, review platforms, Reddit threads, directories. Apollo.io had 19 review-directory profiles and 24,000+ accumulated reviews, and got zero citations on CRM queries: the consensus had filed it under "sales engagement," not CRM. The label decides which shelf you sit on.

And the channel that weighs most in that consensus is one classic SEO barely touched: communities. Reddit appears in Google's results for 81.6% of SaaS buying queries, and a third of the threads cited by AIs are over a year old. Machines trust the forum more than the vendor's blog. One honest caveat: Reddit today aggressively removes comments with brand affiliation; it's best-effort visibility, not a channel you can fill on demand.

Query fan-out: one question, 6-8 hidden searches

Here's how retrieval works inside Google AI Mode and AI Overviews: your question isn't searched as-is. The system unfolds it into multiple parallel sub-searches (the "query fan-out"), retrieves pages for each, and synthesizes. - Source: Google Search Central

Google's own documentation example: for "how to fix a lawn that's full of weeds," the fan-out fires "best herbicides for lawns," "remove weeds without chemicals," "how to prevent weeds in lawn," and so on.

One fan-out capture study (EMGI) logged the pattern: a buying query generates 6 to 8 hidden sub-queries, touching your comparison page, your pricing content, a listicle you appear in, and a Reddit thread about you, all simultaneously. The game went from 1-to-1 to 1-to-many: you don't rank for a query, you get retrieved across a mesh. A brand present on 7 branches of the fan-out beats one that's perfect on 1.

The nuance almost nobody mentions: retrieval doesn't happen evenly across the funnel. In the Reddit citation study (1,486 queries), top-of-funnel questions ("what is a CRM?") got direct citations in only 2.5% of cases; the model already knows what a CRM is and answers from memory. Bottom-of-funnel buying queries: 20.1%. Meaning: AEO effort on definitional content goes straight in the trash. Spend it where the model is forced to look things up.

What stays the same (and nobody tells you)

Four things carry over from SEO unchanged, and if you skip them, no GEO tactic saves you:

1. Crawlability. If an agent can't fetch and parse your page, nothing else exists. And here's an uncomfortable datapoint from Vercel: the major AI crawlers don't render JavaScript. Your Angular SPA with client-side rendered content is invisible to the crawler exactly the way it was to Googlebot without JS.
2. Topical authority. Models retrieve from pages that already rank. Every serious analysis of AI Overview sources finds heavy overlap with the organic top 10. Your ranking is your ticket into the retrieval pool.
3. E-E-A-T. Experience, Expertise, Authoritativeness, Trustworthiness: it was born for Google's human quality raters and turns out to describe exactly what makes a source citable by a machine. Named authors, real credentials, original data, clear sourcing.
4. Clean information architecture. Descriptive headings, logical hierarchy, one topic per page. Helps the crawler, helps the retrieval system, helps the reader. Triple dividend.

If a GEO vendor tells you this no longer matters, ask which of the four they plan to skip. The answer is none.

The format AIs cite

The original Princeton paper didn't just name the field: it measured which edits increase visibility in generated answers. Results, on their 10,000-query benchmark:

• Adding quotes from relevant sources: +41% visibility.
• Adding statistics: +32%.
• Citing sources: +27%.
• Keyword stuffing: 17.7 against a 19.3 unedited baseline. Worse than doing nothing.

In 2026 the mechanism is better understood: generative engines don't index whole pages, they cut content into "chunks" (passages) and retrieve the most relevant segments to synthesize. Your page doesn't compete as a unit; each section competes as an independent snippet. That translates to writing like this: every H2 carrying a complete, self-contained answer, not depending on the previous paragraph to make sense.

On concrete formats, an analysis of the 25 most-cited domains for one B2B client found: ranked lists present in 19 of 25, topic guides in 17, comparisons in 3, case studies in 2, how-tos in 1. - Source: agenticinbound The mechanism is obvious in hindsight: a model answering "best X for Y" needs a source that already contains a set of comparable options. A case study about one client doesn't contain that set, no matter how good it is. In classic SEO your case study ranked and converted; in AI answers, 2 out of 25.

Verified smoke: llms.txt and friends

If in 2026 someone sells you llms.txt as an optimization, you already have the data to decide:

• Ahrefs tracked 137,210 domains with llms.txt: 97% received zero requests for the file in May 2026. - Source: Ahrefs
• SE Ranking (300,000 domains): 10.13% adoption, no measurable effect on citations; their XGBoost model predicted better when the variable was removed.
• Trakkr (37,894 domains, 337,362 citations): association r = -0.065. Noise.
• And Google in writing: Search ignores llms.txt; it neither helps nor hurts.

Three independent studies, same conclusion. Adopting it does no harm (it's a text file), but if your consultant bills it as a service, you know which part of the invoice is smoke.

The bots: who actually reads your site

Blocking "AI" isn't one decision, it's several. AI companies split crawling into user-agents with different purposes:

Two details that surprise: first, blocking the training bot does not remove you from AI search; they're separate pipelines. Second, per OpenAI, ChatGPT Search reflects a robots.txt change in about 24 hours: bot configuration now has one-day latency, not an indexing-cycle latency.

The strategic call depends on your business: traffic-monetized publishers might block training and allow search; consultancies living off being cited as authority, the opposite. The indefensible position is not deciding out of ignorance.

Measure without fooling yourself

The most common GEO measurement mistake: counting sessions. If your reporting only looks at traffic, GEO "is doing nothing" while it's working, because its output is mentions and citations, not clicks.

What gets measured seriously:

• Mentions vs citations, and they're different things. In agenticinbound's 90-day dataset (5,550 AI responses): a US B2B e-commerce client accumulated 22,885 brand mentions but only 4,133 citations (18.1%). Two small European brands in thin-content niches: 51% citation rate. Being in the gap (niches with little published content) nearly triples your rate without changing your authority.
• Share of voice per engine: a fixed panel of real buying prompts, run periodically, measuring how often you're named and cited. It's the only way to have a baseline against which to detect change.
• AI traffic conversion: here the evidence is contested and I'll say it as-is. Semrush reported a visitor from AI search converts 4.4x better than an organic one. A later Search Engine Land study with affiliate data found the opposite: ChatGPT traffic converted worse than Google organic (organic was +13% better) and represented just 0.2% of total sessions (~200x less than Google). - Source: Search Engine Land My read: it depends on site type and intent. Measure yours, don't import anyone's average.

And one hygiene rule: a brand mention without a link is not a citation. When auditing, verify the cited page actually contains the evidence the answer attributes to it.

Final checklist

The actionable list, in order of return:

1. Audit your crawlability for AI bots. robots.txt per bot, and verify what's visible without JavaScript. If your critical content only exists after a client-side fetch, neither Google nor ChatGPT can cite it.
2. Decide your bot policy (training vs search); don't inherit it by default.
3. Identify your bottom-of-funnel queries and concentrate extractable content there: definitions the model already knows; comparisons and "best X for Y" are where it searches live.
4. Turn your best content into self-contained chunks: every section answers something complete, with numbers and cited sources. Ranked lists and topic guides win citations; case studies don't.
5. Work the consensus off your website: reviews, directories, communities, editorial coverage. That's where the label models file you under gets decided. Remember Apollo.io: 24,000 reviews and zero citations for being on the wrong shelf.
6. Build your prompt panel and measure share of voice monthly. Mentions and citations separately.
7. Ignore llms.txt without guilt and distrust any service selling "GEO" as separate from SEO: Google itself says it's the same thing, and the four pillars (crawl, authority, E-E-A-T, structure) haven't changed.

(In fact, this post is written that way on purpose: every section answers something complete without leaning on the previous one. If this article shows up cited in your next ChatGPT answer, it's because the method works. If you want to see how I apply this to other topics, check out my blog.)

And one last thing, the most important: none of this replaces SEO. It extends it. Your ranking is still your ticket into the retrieval pool, your authority is still the strongest signal, and the old metrics still matter. The difference is the prize now has two outputs: the click and the citation. Optimizing only for the first one, in 2026, means leaving half the field unplayed.

Have you measured how often ChatGPT cites you for your niche's queries? If the answer is "no idea," that's your next weekend.