How to Get ChatGPT to Recommend Your Product: A Practical Guide to Answer Engine Optimization

Smith got into SEO in 2007 doing programmatic and commerce SEO for shopping comparison sites, when mass auto-generated landing pages worked. In his view the biggest change was Google's Panda and related algorithms, which turned that kind of SEO into spam. He ranks AEO as the second-biggest, because it is related to search but is a summarization of search with new inputs. Host Lenny Rachitsky notes that many people see AEO as making everything different, while Smith sees continuity.

On naming, he treats AEO and GEO as the same thing: how to show up in LLMs as an answer. He prefers "answer" because "generative" could also mean images and video. Rachitsky says ChatGPT now sends more traffic to his newsletter than Twitter, and Smith says clients saw a jump in January from both higher adoption and answers becoming more clickable through maps, shopping carousels and cards.

Asked about ChatGPT head Nick Turley's advice to just write great content, Smith partly agrees. Anything can be optimized once you understand the rules, but he reads the message as a warning not to spam. Google eventually wiped out whole categories of auto-generated pages, and he expects ChatGPT teams to do the same to anything spammy.

For Webflow, Smith's team did traditional SEO such as landing pages for high-volume keywords like "best no-code website designer", which brings AEO benefit for free. He says everything that works in SEO works in AEO, but the head and the tail differ. At Google, Webflow's blue link at number one would win. In an LLM, a number-one citation does not win the answer because the model summarizes many sources, so you need to be mentioned across many.

The work that did best for Webflow was straight SEO, YouTube and Vimeo videos, and Reddit optimization. This matters for startups: he normally tells them not to bother with SEO early, since domain authority takes years. A new citation, such as a Reddit thread, a video, or a blog post about a newly launched YC company, can put a brand in answers tomorrow.

The tail is also bigger. He cites Perplexity as saying the average prompt is around 25 words versus about six in Google, with many follow-up questions and questions never searched before. Long-tail SEO of one page per keyword stopped working, but in chat the long tail is back. He has seen early-stage companies with obscure products, such as an AI-enabled payment processing API, show up simply by answering questions nobody had answered.

Smith says LLM leads are far more valuable: Webflow saw a 6x conversion-rate gap between LLM traffic and Google Search traffic. He thinks it comes from users being primed by a multi-step conversation that narrows their intent. He splits the work into onsite, which is traditional SEO plus answers to detailed product questions about use cases, features and integrations, and offsite, which means showing up in citations such as video, UGC like Reddit and Quora, affiliates and blogs.

Reddit is the most-asked-about piece. The growth-hacker instinct is hundreds of fake accounts auto-posting and upvoting their own comments, but those accounts get banned and comments deleted. The approach that works at Webflow is for a few employees to find threads that are already cited, say their name and employer, and give genuinely useful information. In his words you don't need 10,000 comments, and even five can be great.

Rachitsky points out that Deel grew early the same way on Reddit. Smith adds that ChatGPT itself acts as a police: its search algorithm selects which citations it trusts, and people there tune it. Google does the same with Reddit, Twitter and Quora, so if the content were bad, they would stop ranking it.

Smith distinguishes the core model from retrieval-augmented generation, which is essentially search followed by summarization. Influencing the core model would be extremely hard and slow, perhaps requiring a million pages saying a brand is best for something, so he focuses on the controllable RAG side. An LLM is also unlikely to name your product if it appeared nowhere in retrieval.

He lists four fundamentals: LLM plus RAG, topics, question research, and citation optimization. Each page should target a topic that bundles hundreds or thousands of related questions, and answering more follow-ups, especially ones others skip, raises your chance of appearing. Because ChatGPT gives no search-volume data, he suggests converting your keywords into questions and mining sales calls, customer support and Reddit for the questions people really ask.

He describes Dotdash Meredith, owner of Good Housekeeping, Allrecipes and Investopedia, as probably the most successful SEO company ever and the most-cited in LLMs. He says much SEO content is non-experts rewriting each other, and notes that about one in 20 landing pages drives roughly 85% of traffic. To fix this, search engines could reward information gain, meaning you said something new, and penalize being typical. Rachitsky notes that the same heuristic makes for good newsletters and podcasts.

Smith's plan starts with picking the questions you want to win. Take your own and competitors' paid-search terms, since those are money terms, and ask ChatGPT to turn them into questions. Then put them into an answer tracker, see who is cited, and build your own landing pages of the type that appears most, whether listicle, category page, article or tool page. Each page should answer all the likely follow-up questions.

Offsite, he treats each citation group differently. Paying an affiliate such as Forbes is expensive but easy and controllable. YouTube and Vimeo videos are easy because no community polices them, and there is an opening in unglamorous, high-value B2B topics that few people make videos about. Reddit is easy but less certain because the community might reject your answer.

He insists on experiments, because he says most SEO and AEO best practices are repeated without analysis and are wrong. Track questions for a couple of weeks, make changes on a test group, and compare against a control group. For staffing, the SEO team can take on onsite work, while citation work like video and Reddit may need a community or marketing generalist.

Smith explains that answer tracking is like keyword tracking, but ChatGPT's answers are a weighted random sample of a distribution, so the same question gives different answers each run. Different phrasings and different surfaces such as Perplexity, Gemini, ChatGPT and Meta AI also vary. You therefore need to measure a share of voice and average rank across repeated runs and variants. His advice is to pick the cheapest of the roughly 60 tools that does what you need.

The underlying algorithms are similar, but results differ. In a study of thousands of questions, citation overlap with Google Search was around 35% for ChatGPT and about 70% for Perplexity. On whether to worry beyond ChatGPT, he compares it with AOL in 1999: nobody knows who will win, so expect around three surfaces that matter, rather than ten.

For B2B, the cited sites differ from consumer ones: TechRadar might show up often, whereas commerce skews to Glamour and Cosmopolitan and marketplaces to Eater, Yelp and TripAdvisor. Most B2B answers are not clickable, so last-touch referral traffic undercounts impact. Smith recommends tracking whether you appear in answers and asking customers post-conversion how they heard about you, since B2B buyers go through around 50 touchpoints.

Commerce answers show shoppable cards with multiple sellers, so schema and reviews matter and referral traffic is a decent measure. For early-stage companies he says to do only citation work and long-tail answers and skip mid-funnel SEO. He also warns that users often open a new tab and type your brand into Google or your domain directly, so some LLM traffic is mislabeled as branded or direct.

On whether publishers should let AI engines ingest their content, Smith echoes Brian Balfour that you are playing the game whether you choose to or not. Blocking indexing hands the slot to competitors. If you object to training, you can use robots.txt to disallow training bots while allowing indexing bots.

Smith describes an experiment design: take about 200 questions, leave 100 untouched as a control, and apply an intervention to the rest, such as Reddit comments, a video, or a paid affiliate mention. Track a couple of weeks before and after. A control group is essential because answers vary a lot on their own and LLM usage is rising.

He stresses reproducibility, drawing on his academic research background. In SEO it is common to credit the wrong cause and assume it works forever, so he suggests repeating studies and reading others' results, trusting a tactic only after it works many times. Most work in SEO and AEO is wasted, he says, and only testing separates what works.

On how big the channel is, he says it is smaller than search and paid, but substantial: Webflow gets about 8% of signups from LLMs, which makes it one of its top channels.

Smith says misinformation is unusually high. Every couple of years articles claim Google Search is dying because of a new thing, whether TikTok, Instagram, Facebook or now AI Overviews, but each is a new surface that adds to Google rather than taking from it: "Google's slice of the pie stays the same. The pie gets bigger." He notes that Google's VP of search said traffic sent to publishers was up slightly.

He also sees expensive tools doing commodity tasks, with people paying huge amounts for what is essentially keyword tracking because the space seems mysterious. The channel's growth curve is also unusual: interest was low after a Reforge webinar last June, then skyrocketed in January. He says many claim SEO and AEO are completely different to sell tools, when there is considerable overlap and the core technology is similar.

On why January was an inflection point, he points to both growing LLM adoption and rising clickability, especially for commerce, local and hotels, where rich modules let people click through.

When ChatGPT launched, Smith predicted spam, because he had made similar scraped, auto-generated pages in 2007 and watched Google wipe them out. He distinguishes automated content with no human in the loop from AI-assisted content, which he considers the clear future. To test it, his team took thousands of searches and questions, collected Google results and ChatGPT citations, and ran them through Surfer SEO's AI detector, which they validated on known AI articles and found about an 8% false-positive rate on pre-ChatGPT pages.

They found only roughly 10% to 12% of results in Google and ChatGPT were AI-generated, and the correlation analysis agreed that purely AI content does not win. They also found, in a Common Crawl sample of 100,000 URLs over five years, that there is now more AI-generated content on the web than human-written content. If it did work, everyone would do it, and Google or ChatGPT would become a search engine for AI output, which is exactly what led Google to remove the shopping comparison sites.

He then describes a loop in which AI summarizes its own derivatives, linking it to the model collapse paper on training. In retrieval, the LLM relies on a wisdom of the crowd, and an infinite loop of derivatives would converge on a single opinion, such as "vanilla" as the only ice cream flavor. Rachitsky says he fears a world of AI trained on AI, and asks how the platforms will avoid perverse incentives.

Smith expects LLMs and search to converge into a single experience, with Google adding AI Overviews and LLMs adding maps and shopping carousels. He advises asking what 2007 Ethan would do wrong and not doing it. He also imagines autonomous agents that remember everything about you and make decisions like a good executive assistant, for example planning a trip without intervention, which would be another thing to optimize for.

His most exciting tip is help-center optimization. Since chat users ask follow-ups like whether a product supports a use case or integration, help centers can answer them. He suggests moving the help center from a subdomain to a subdirectory, cross-linking pages heavily, and writing content for the tail, such as an obscure integration question he had about meeting transcription tools and Looker. Opening the help center to community questions can fill in the tail, and you might be the only citation for those questions.

Smith recommends Emotional Intelligence, Robert Cialdini's work on persuasion, Thinking, Fast and Slow, and How to Measure Anything, arguing that psychology and behavioral economics give growth people frameworks. He likes aggressive sports documentaries like The Last Dance, and UFC, balanced by climbing films with Alex Honnold, which he links to his own mix of intensity and zen craftsmanship. His motto draws on Outliers: be sufficiently smart, practice in a focused way, and practice a lot.

His favorite SEO win is ranking MasterClass for "butter lettuce" against stronger sites like Allrecipes and Martha Stewart. He publishes studies on LinkedIn and on his company's blog, called The 5%, referring to the idea that 5% of landing pages drive almost all impact. He asks listeners to share AEO test results and to leave long, thoughtful comments on his LinkedIn posts.