Notes — Ethan Smith on Answer Engine Optimisation, LLM Search, and the Citation Playbook

Four questions [Adler frame]

Q1 — What is it about? A practitioner guide to AEO (Answer Engine Optimisation) — the discipline of optimising your product’s appearance in LLM-generated answers (ChatGPT, Perplexity, Gemini, Claude). Ethan Smith is the CEO of Graphite and has 18 years in SEO from programmatic commerce (NexTag, Shopping.com, 2007) through the present. His core claim: AEO is the second-biggest shift in search since Google introduced spam-prevention algorithms (Panda et al.); everything in SEO also works in AEO, but AEO has additional levers. Illustrated with Webflow client results.

Q2 — How is it argued? Through practitioner evidence and controlled research, which is unusual for this topic. Key data points: Webflow saw a 6× conversion rate from LLM traffic vs. Google traffic; LLMs now account for 8% of Webflow signups; a proprietary AI content study (thousands of URLs, validated AI detector, pre-ChatGPT false-positive baseline of ~8%) found only 10–12% of Google and ChatGPT citation content is AI-generated, disconfirming the claim that AI content works at scale; a model-collapse study showing that derivatives of derivatives converge to a single opinion. The argument’s structure is: AEO is RAG (search + summarisation), so citation count — not rank — determines outcomes; therefore the playbook differs from SEO mainly at the head and in the citation optimisation layer.

Q3 — Is it true? Largely credible. Smith’s AEO study methodology is notably rigorous — he validated the AI detector against a pre-ChatGPT baseline, which is the right control. The 6× conversion rate data is from a single client (Webflow) and may not generalise; however the directional claim (intent is higher because chat queries carry more context) is well-supported. The head-vs-tail distinction (25-word avg vs. 6-word avg) references Perplexity data. Google not dying: supported by Google VP of Search statement cited in the episode. Model collapse claim references an actual academic paper. No contested claims.

Q4 — What of it? Three audiences: (1) growth/marketing people at B2B SaaS who have not yet optimised for LLMs — the 7-step playbook is directly actionable; (2) early-stage founders — for the first time in ~15 years, new companies can compete at head queries against incumbents, because citation count (not domain authority) determines ranking; (3) content strategists — the AI-generated content finding is important: 100% automated AI content does not rank, AI-assisted content does. Help centre optimisation is the most counterintuitive finding: a product’s help docs are now a primary AEO asset because they answer tail questions nobody else is answering.

Glossary

AEO / GEO — Answer Engine Optimisation / Generative Engine Optimisation; terms for the same discipline. Smith prefers AEO because “answer” is more specific than “generative.” Operationally: how do I show up in LLM-generated answers?

RAG — Retrieval-Augmented Generation; the architecture underlying most LLM search features. The model performs a web search, retrieves citations, then summarises. AEO operates on this layer, not on the core model weights. Controlling for this distinction matters: influencing the core model (weight-level) is impractical; influencing RAG citations is tractable.

Citation optimisation — getting your product mentioned across the sources that LLMs trust and retrieve: your own site, YouTube/Vimeo videos, UGC (Reddit, Quora), tier-one affiliates (Dotdash Meredith: Good Housekeeping, Allrecipes, Investopedia), tier-two affiliates, blogs. LLMs summarise from all of these simultaneously.

Share of voice — the percentage of times your brand appears in LLM answers for a set of target questions. The correct AEO metric (replaces rank in traditional SEO) because LLM answers are probabilistic — the same question returns different answers each run.

Head vs. tail in AEO — Head: competitive “best X” queries where citation count, not citation rank, determines whether your brand appears first. Differs from Google where rank 1 = win. Tail: queries averaging ~25 words (vs. ~6 in Google Search), many of which have never been searched before; early-stage companies can win the tail immediately because there is no incumbent domain authority advantage.

Information gain — heuristic for content quality: does this page say something that existing sources do not? Google and LLMs both theoretically favour it, though Smith notes EEAT weighting is inconsistent in practice. Primary defence against becoming a rewritten derivative.

Model collapse — academic paper concept: if LLM-generated content is fed back into RAG citations, and those derivatives are used to generate new derivatives, the system converges towards a single opinion, losing the diversity of the “wisdom of the crowd.” Smith applied this to the RAG layer specifically.

Help centre optimisation — treating a product’s help docs as an AEO asset. Move to a subdirectory (not subdomain), add internal cross-links, and fill the tail with use-case/integration questions customers actually ask on sales calls and support tickets.

AEO seven-step playbook

Smith’s operational framework for winning at AEO:

1. Identify target questions. Take paid search keyword data (yours + competitors’). Feed into ChatGPT: “Turn these into questions.” Supplements with: sales call recordings, customer support tickets, Reddit threads about your product.
2. Set up answer tracking. Choose any of ~60 answer-tracking tools (analogous to keyword trackers). Track share of voice across question variants and surfaces (ChatGPT, Perplexity, Gemini, Meta AI). Prioritise 2–3 surfaces with most traffic.
3. Audit current citations. For your target questions, identify which URLs LLMs are citing — categorise by type (own site, video, UGC, affiliates, blogs). These define your competition and your opportunity gaps.
4. Create on-site content. Match page type to what is currently ranking (listicle, category page, article, tool page). Answer all follow-up questions exhaustively — more questions answered = more likely to be cited. Help centre docs are an underutilised asset.
5. Execute off-site citation strategy by group:
   • Affiliates (Dotdash, TechRadar, etc.): paid placement if budget allows — expensive but controllable
   • YouTube/Vimeo: make product videos; B2B vertical has very low competition in video for specific topics
   • Reddit: create an account with your real identity, say where you work, give a genuinely useful answer in relevant threads. Do not automate; community moderation will remove fake accounts. Even five high-quality comments can move share of voice.
6. Design and run experiments. Split 200 questions: 100 control (no changes), 100 test. Make one intervention. Wait 2–3 weeks. Share of voice went up in test but not control = the intervention worked. Reproduce before treating it as settled.
7. Build the team. SEO team owns on-site + answer tracking. Marketing/community generalist owns YouTube and Reddit strategy. Most SEO agencies can do steps 1–4; steps 5–6 require different skills.

The head vs. tail distinction — mechanics

The most important structural difference from traditional SEO:

Head queries (e.g. “best website builder”):

• Google: being #1 in blue links = winning; domain authority determines rank
• LLM: summarises multiple citations; the product mentioned most often across citations appears first — not the product whose URL is ranked #1
• Implication: a new company with no domain authority can win by being mentioned in 10 Reddit threads, 3 YouTube videos, and 2 affiliate articles before a larger competitor with one highly-ranked page

Tail queries (e.g. “which meeting transcription tool has a Zapier integration to Looker?”):

• Average query length ~25 words (Perplexity data) vs. ~6 words for Google
• Many such queries have never been searched before — no existing citation exists
• First publisher to answer the question wins by default
• Help centre docs are the primary vehicle; community forums are a secondary vehicle

Early-stage implication: Smith explicitly reverses his own SEO advice for early-stage companies. Old advice: don’t do SEO until Series A/B (no domain authority). AEO advice: definitely do AEO from day one — win the tail immediately, get cited as soon as possible.

AI content study — methodology and findings

Smith’s research team conducted a controlled study:

1. Retrieved thousands of Google Search results and ChatGPT citation URLs for target queries
2. Applied Surfer SEO’s AI detector
3. Validated detector accuracy: (a) generated thousands of known AI articles → high true-positive rate; (b) sampled 100,000 URLs from Common Crawl pre-ChatGPT launch → false-positive rate ~8%
4. Finding: ~10–12% of cited content in both Google and ChatGPT is AI-generated. 90% is human-written.
5. Correlation analysis: AI-generated content does not rank/get cited; AI-assisted content does.
6. Bonus finding: more AI-generated content exists on the open web than human-generated content (by volume, post-2023) — but it does not surface in search results.

Model collapse extension: Smith’s team simulated feeding derivatives of derivatives into RAG. Result: diversity of answers decreases, converges on a single response. Wisdom-of-the-crowd effect requires source diversity; derivative loops destroy it.

Practical implication: 100% AI-generated content without human editing is a waste of resources and may harm brand reputation via association with spam. AI-assisted content (human-directed, AI-accelerated) is the correct approach.

B2B vs. commerce vs. early-stage differences

Related

• Ethan Smith on Answer Engine Optimisation, LLM Search, and the Citation Playbook — transcript page
• Answer Engine Optimisation — derived concept page
• Ethan Smith — speaker page
• Eli Schwartz on Product-Led SEO and the AI Overviews Shift — adjacent episode: SEO transition, AI Overviews
• Nick Turley on ChatGPT — Smith explicitly disagrees with Turley’s “just write great content” AEO advice