Article details
Recap from Ethan's Reforge Event: Research-Backed Strategies for Maximum AEO and SEO Impact
Most product and marketing teams have spent years learning how to show up in Google. They know keywords, backlinks, authority, page structure. They have dashboards. They have playbooks.
What they don't have is a playbook for the channel that's growing fastest right now: AI answers. That's what Ethan Smith — CEO of Graphite, a research-driven SEO and AEO agency — laid out in a Reforge session: an end-to-end workflow for building an answer engine optimization strategy from scratch.
The session covered five things in sequence: how to find the prompts you want to win, how to track them, how to build on-site strategies, how to build off-site strategies, and how to measure what's actually working.
Why AI is not just a longer search
Before you can build an AEO strategy, you need to understand what makes AI different from search — because if you treat them the same, you'll optimize for the wrong thing.
Search queries are short. In practice, the long-tail keyword distribution you've seen in SEO presentations is mostly theoretical. Real search behavior clusters around two to four words. Users type something short, scan results, click, and move on. Search is optimized for speed.
AI prompts are the inverse. The vast majority of prompts are ten words or more. Users are adding context — their situation, their constraints, their persona, their comparison set. "What's the best CRM?" becomes "What CRM is designed to unify customer data across dozens of systems, integrates with our existing data warehouse, and doesn't require heavy IT involvement to administer?"
This has an important implication for the narrative that AI is replacing search: it isn't. Search usage is up, not down. AI is a net new behavior, not a substitution. The channels serve different moments in a user's journey, and the prompt length difference tells you why — AI handles the complex, context-heavy questions that search was never great at answering.
Finding the prompts you want to win
Unlike search, there's no Google Search Console for AI. There's no authoritative list of prompts that people are typing. So the approach is to triangulate from multiple imperfect sources — overlapping circles in a Venn diagram, each of which gives you a window into a different part of real prompt behavior.
Search data gets you the head of the distribution. The method Ethan recommends is not to use related keyword suggestions, which tend to be noisy. Instead, take a competitor's URL that ranks for the topic you care about, look at what keywords that specific page is ranking for, and use an LLM to group those keywords into sub-themes and transform the parent of each group into a question. Three hundred variations become fifteen prompts, sorted by search volume.
Reddit gets you the tail. Search data tells you what people search for in short form, but prompts live in the long tail — and Reddit is where users ask detailed, context-rich questions in a format that actually maps to how they'd prompt an AI. Use Gemini rather than Claude for this, because Google has a larger Reddit corpus. Ask it to find threads about your topic, identify the most common questions, and sort by upvote count as a proxy for volume.
AI brainstorming gets you an additional signal. It's not primary source data, but a large language model trained on billions of documents has absorbed an enormous amount of implicit signal about what people ask. Asking an LLM to generate likely prompts for your topic is a reasonable supplementary method — treat it as one more overlapping circle, not a primary source.
Bing Webmaster Tools is the most underused asset in this stack. It's the only first-party data source that shows you the actual grounding queries being used when your content surfaces in Copilot. Microsoft simply chose to expose this data; Google hasn't yet, though it presumably will. Even accounting for Copilot's smaller market share, first-party data is always more valuable than inference.
Internal data often gets overlooked, but for B2B companies especially, your sales calls are full of real prompts. If you're using Fathom or a similar transcription tool, you can pull your call transcripts into a data warehouse and ask an LLM to surface the questions people are actually asking before they buy. These are the questions your prospects are probably also typing into AI assistants.
Tracking: what you're actually measuring
Once you have a set of prompts, you need a way to track your visibility across them. Most teams are already using tracking tools, but there are a few things worth doing differently.
First: don't just look at the top domains that appear in AI answers. Reddit, Wikipedia, and YouTube are almost always in the top domains — but they're not the top categories. When you look at the distribution by category, the picture is different: corporate sites (your competitors, your own site) and editorial sites (blogs, affiliate content, review sites) are usually the dominant categories. Build your strategy around categories, not individual sites.
Second: separate branded prompts from non-branded ones. If you include "is [your company] a good product?" in your overall visibility metric, of course your visibility looks high. Branded and non-branded prompts tell you completely different things about your position in the market and require different strategies.
Third: remember that AI is probabilistic. Ask the same prompt ten times and you'll get ten different responses. That's not a bug — it's how the technology works. It also means you can't draw conclusions from a single response. Ask each prompt five to ten times to understand the distribution of answers, not just a point-in-time snapshot. As you run more iterations, the variance decreases and the distribution becomes meaningful.
On-site strategy: build consensus
The core insight of on-site AEO is counterintuitive to anyone who has internalized standard SEO practice. In SEO, you want one authoritative page per topic. Cannibalization is a real risk. You don't want six pages on the same subject.
In AEO, the more the better. AI synthesizes consensus — it's reading across many sources and constructing an answer that reflects what the majority of credible content says. If six of your pages explain different facets of why your product does a particular thing, AI is more likely to represent you accurately and prominently than if you have one. You are building the narrative that the model assembles.
Looking at what actually surfaces in AI citations reveals the kinds of pages that work. For B2B SaaS, it's primarily product content (feature pages, use case pages), help center content, and forums. For commerce, it tends to be collection pages and product pages. The pattern isn't about page authority — it's about what the page says. Pages that clearly, specifically answer "yes, we do this thing with these specific attributes" are what AI reaches for.
This has an important practical implication: your product pages probably have gaps. When AI surfaces a prompt like "what CRM handles unified customer data across dozens of systems for non-technical teams," it's looking for pages that explicitly claim those capabilities. Most product pages explain the primary use case. The secondary and tertiary use cases, the long-tail attribute combinations, the integrations, the personas — these often aren't written down anywhere. Writing them down is a significant part of an on-site AEO strategy.
Off-site strategy: owned and earned
Off-site in SEO means backlinks. Off-site in AEO means something broader — and the distinction between owned and earned offsite is more important than people realize.
Earned offsite is the content other people produce about you: Reddit threads, blog comparisons, review sites, media coverage. You don't control this content directly, but you can influence it. The approach that actually works on Reddit is engagement, not manipulation. If someone writes that you're overhyped, go respond — not to argue, not to say they're wrong, but to engage authentically and thoughtfully with the criticism. Ethan did this when a Reddit thread appeared questioning whether he was overhyped as an AEO expert. He wrote a dispassionate response that engaged with the arguments and ended up doing an AMA with the subreddit moderators because his responses were genuine enough to earn their trust. Reddit AMAs are another option — especially for addressing specific misunderstandings about your product. One client, a dog food brand, used an AMA with a board-certified vet to correct factual mischaracterizations about whether their product was evidence-based, and measured a clear shift in brand sentiment afterward.
Owned offsite is where most companies leave significant value on the table. Your GitHub profile, LinkedIn page, Instagram, TikTok, app store description, YouTube channel — you control what all of these say about you. And they do appear in AI citations. In the N8N case study, N8N's own GitHub profile description was a cited source in answers about what N8N is known for. Their pricing page was cited. The description they wrote about themselves on their own platforms became part of the narrative AI tells.
This means your profile descriptions are now brand assets that should be treated with the same intentionality as homepage copy. Whatever you want to be known for — and especially whatever you're not currently being credited for — should be reflected in the descriptions you control across every platform where you have a presence.
One clarifying exercise: ask an AI directly what it knows about your brand. What is your product known for? How would it compare you to competitors? What's missing? For N8N, this analysis revealed they were scoring well on flexibility and self-hosting but poorly on governance, observability, and enterprise trust signals — capabilities they had but weren't communicating. That gap between capability and narrative is the content gap, and closing it is the work of both on-site and off-site AEO.
Measuring impact
Most AI answers don't produce a click. The user gets their answer, maybe opens a new tab, maybe types your brand name into Google, maybe comes back a month later. Last-touch attribution will not tell you any of this happened. That's not a data problem to solve later — it's a fundamental limitation of attributing AI influence with traditional measurement.
The best measurement approach is a self-reported survey. When someone arrives at your site, ask how they heard about you. When they select AI, follow up: which assistant? And then: what prompt did you type? This is real first-party data — not inferred, not modeled, not probabilistic. Ethan implemented this at Graphite and found that self-reported AI attribution was ten times higher than last-touch attribution from direct or referral traffic, which tells you how much the standard approach was undercounting.
The "what prompt did you type?" question is particularly valuable. It closes a loop that's otherwise invisible: you now know which specific prompts are actually driving your conversions, you can verify that your AEO work is targeting the right things, and you can take those real prompts and feed them back into your tracking and optimization workflow.
For more sophisticated measurement, mixed media modeling is the right framework — building a model that accounts for multiple channels' contributions to a conversion rather than attributing it to the last thing that happened. It's complex to implement, but it's the right direction for any team serious about understanding how AI sits in the larger customer journey.
Where to start
If you're doing SEO and have reasonable authority (roughly Series A or later), the right initial split is approximately two-thirds SEO and one-third AEO. SEO is a larger and better-understood channel, and most of your SEO work transfers to AEO for free. The incremental AEO work is primarily product content, brand content, and owned offsite — getting your profiles and descriptions right — rather than the more expensive and harder-to-control work of off-site earned coverage.
If you're earlier stage and in a niche category, AEO can actually deliver impact sooner than SEO. You don't need domain authority to build consensus around a specific new capability — you just need the right content. For a brand-new product, start with product content that explains your specific use cases and personas, get your owned offsite profiles in order, and set up Bing Webmaster Tools so you have at least some first-party visibility data.
Whatever stage you're at: add the prompt question to your onboarding survey. It's the simplest way to start gathering real data on what AI is doing for you, and it costs almost nothing to implement.
Most of the industry is still running a 2024 playbook in a world that has already moved. The teams that figure out the prompt layer first will have compounding advantages that are increasingly hard to overcome — because once an AI's base model has associated your brand with a capability, it takes a lot of consistent, credible content to change that association. The time to build that position is before your competitors do.
The one thing to test and measure
Ethan closed with a reminder worth repeating: many AEO best practices that circulate online haven't been verified with real data. The research is often methodologically weak, and the channel is young enough that what worked six months ago may not work today.
Test your own stuff. Generate hypotheses from sessions like this one, from your own prompt tracking, from your own survey data. Run experiments. Build consensus around the things that actually matter to your customers and measure whether your visibility changes. The teams that will win in AEO are not the ones that follow playbooks most faithfully — they're the ones that treat it as a research problem and keep updating their priors.
Watch the full recording here
Find out more about Ethan's upcoming cohort of Mastering AEO