Ravish Agrawal and I get along for a simple reason: we're both allergic to bullshit. Neither of us trusts a number we didn't wire ourselves. Neither of us has patience for a visibility chart that goes up while signups stay flat.

Ravish is Head of Growth Marketing at Gamma, which crossed $100M in ARR profitably with about 50 people. I built the revenue marketing function at Webflow and now do this work with companies through StackedGTM. When someone asks me what AEO looks like when it is run as a signup channel, Gamma is a company I point to.

So I sent him the ten questions I'd want answered if I were starting from zero, and I asked for specifics. He gave them. He also says plainly where Gamma hasn't run the test or doesn't know the answer yet, which is the part I trust most.

The questions are mine and the answers are his, published as he wrote them.

Josh: Let's kill the biggest myth first. Most people think you get cited by publishing more content on your own domain. You and I both know the models pull from a fan-out of sub-queries and lean hard on third-party surfaces: Reddit threads, YouTube, G2, comparison sites, other people's blogs. What percentage of Gamma's citations actually come from gamma.app versus places you don't control, and what did you do to influence the places you don't control? That's the part nobody publishes.

Ravish: Most of our citations come from outside gamma.app. Gamma’s first hundred million users came from video, so there was relatively little written about us when people started asking ChatGPT about presentation tools. We started with the sources the models were already using: Reddit, then comparison articles and listicles cited for the questions we cared about. We asked to be included, or moved higher if we were already listed, because the models seemed to pick up the first few names. YouTube became more relevant as we noticed the Google engines citing it more.

Most of this was outreach. Citation trackers helped us identify which pages to contact.

Josh: You've said you write content for agents, not humans, and half of marketing heard that as heresy. I heard "entity resolution." Show me the page. What does a Gamma page built for retrieval look like that a brand team would have killed in review, and how do you know the agent read it and didn't just find you through a Reddit thread you had nothing to do with?

Ravish: There’s a guides section on gamma.app linked from the footer. We deliberately kept it out of the main navigation. The pages answer the question in the first sentence and cover similar topics in different ways. The subheadings follow the smaller questions a model might search for when putting together an answer. We also include trade-offs and cases where another tool is better suited to the job. Pages that are entirely positive about Gamma seem to get picked up less often.

The guides are repetitive and don’t always read like our usual marketing. We wrote them primarily for retrieval. To check whether they’re being used, we look at crawler fetches, citations for the tracked prompt and visits that lead to signups with an AI referrer. We look for agreement across those signals rather than drawing a conclusion from one alone.

Josh: AI search is a top three signup channel at Gamma converting 20% higher. Walk me through the plumbing. Referrer capture, prompt-level attribution, the self-reported "I asked ChatGPT" cohort, whatever you actually trust. And then the harder one: have you run a holdout? Do you know the incremental number or just the attributed one? Because the whole industry is reporting attributed and calling it incremental.

Ravish: Adding ‘an AI assistant’ to the signup survey made the biggest difference to measurement. People were already finding us that way, but the channel was sitting inside Direct. The new option made it visible without a change in the traffic itself. That self-reported group is still the main measure I use to run the channel.

Referrer data from ChatGPT and other assistants gives us a lower bound, since some people read an answer and then click the URL. Citation tracking helps us understand visibility, but I’ve seen those charts rise without any movement in signups. Crawler logs are useful for checking which engine fetched which page and investigating problems.

We haven’t run a clean holdout. It’s difficult to isolate organic mentions by market when the same sources are available across countries. We have self-reported attribution and a few instances where an external change affected the channel and we could observe what happened to signups. I describe that as attributed acquisition, not proven incremental growth.

Josh: Every model is a different search engine now. ChatGPT, Google AI Mode, Perplexity, Claude, Gemini. They fan out differently, they trust different sources, they cite differently. Which one drives real revenue for Gamma, which one looks great on a dashboard and does nothing, and which one are you deliberately ignoring? Name them. Vague answers here are why the agency market is a mess.

Ravish: ChatGPT gets most of our attention because that’s where most people are. It uses a lot of sources outside our site: forums, threads and other people’s articles. The Google engines feel closer to SEO, where relevance and search rankings still matter. A fair amount of our SEO work is about appearing in Gemini.

Perplexity picks up content quickly, but I wouldn’t invest in it separately given the overlap we see with Google. Claude is a smaller source of discovery for us. I’m more interested in it as a place where someone can use Gamma through an assistant.

The engines don’t respond to the same changes. Someone I know at another company saw two engines drop their microsite after a Google penalty, while ChatGPT barely changed. That’s why I’d test each engine rather than assume one approach covers all of them.

Josh: Answer share isn't page one. In AI answers there's often one recommendation and the rest is noise. I track this as answer capture rate: how often you're the first name, not just a name. What's Gamma's first-recommendation share on your ten most valuable prompts, how much has it moved in 12 months, and what was the single change that moved it most? I want the thing that surprises people, not "we improved our docs."

Ravish: We didn’t work on position for the first several months. We started by getting mentioned, then improving our presence in sources that already cited us. Position came later, once Gamma was regularly among the first few names. The articles and threads supplying those recommendations were a large part of that work.

The biggest change was the format of the content. We moved from posts praising one tool to comparisons explaining what happened when someone tried several. Those seemed to give the models more useful material for a recommendation.

We’ve also seen cases where being first on a list didn’t produce more explicit citations. I don’t fully understand why, so I wouldn’t treat that as a rule. I’d rather have Gamma represented in a broad set of relevant sources than rely on a few pages we control.

Josh: Citations decay. I've watched companies own an answer in March and lose it by June with nothing changed on their end, because the model re-weighted, a competitor got mentioned in the right thread, or a fan-out query shifted. How do you defend answer share once you have it? What's your monitoring cadence, what's the alert, and who on a 50-person team owns it when it drops at 2am?

Ravish: I expected citations to last more like backlinks. When I looked at the data by URL, many pages peaked quickly and faded after a few weeks. Product pages seemed to last longer than blog posts, and our own pages held up better than pages written about us. That’s why outreach, Reddit work and guide updates need regular attention. We review them weekly.

I pay more attention to a change in self-reported AI signups than a visibility score alone. If signups fall while citations stay steady, I’d investigate other possible causes, including the product and pricing. If citations fall but signups don’t, I’d keep watching before making a change.

Most of the time, the changes in the model and the way it retrieves have positively affected us, and our traffic has gone anywhere from 30% to 7x during these changes. We don’t have someone responding to citation changes at 2 a.m. A shift in the sources a model uses generally takes weeks to address. A weekly review gives us time to see whether a change persists and decide what to do.

Josh: Fair objection: Gamma's product creates the exact artifacts people search for. Cheat code. So strip it away. You're Head of Growth at a mid-market CRM tomorrow, no shareable output, no viral loop. First AEO move, and the first thing you refuse to do that every agency would pitch you on day one. I'll go first: I'd refuse to touch llms.txt until I'd fixed my proof density.

Ravish: First, I’d add ‘an AI assistant’ to the signup survey and separate AI referrals in analytics. At the companies I’ve looked at, some of that acquisition was already happening without being identified.

Then I’d try a small set of questions a real CRM buyer would ask and inspect the answers. If the company is absent, I’d look at the cited sources and why it isn’t represented. If it appears but is rarely recommended first, I’d look at how those sources describe it. CRM buyers already have plenty of reviews, migration stories and forum discussions to draw on. We had much less written material about Gamma when we started.

I’d refuse a proposal that started with a dashboard and never explained how it would connect to signups. I also wouldn’t lead with llms.txt. We created one and didn’t see crawlers fetching it. And I wouldn’t create a separate team for SEO and AEO; there’s enough overlap that I’d keep the work together.

Josh: Ads are coming to ChatGPT and Google AI Mode already blends them into the answer. When the paid layer lands, does organic answer share get more valuable or less? My take is the first-recommendation slot becomes the most expensive real estate on the internet and organic ownership of it is the only durable hedge. Agree, or tell me why the paid layer eats the organic answer the way it ate the Google SERP.

Ravish: I think a recommendation can become more valuable even as it sends fewer clicks. We’ve seen organic clicks fall while rankings stayed steady on Google as the results page changed. Better content alone doesn’t necessarily solve that.

Paid placements could also create the same problem as brand search ads: paying to reach people who were already looking for you. I’d want to know whether those placements bring additional customers.

I also think there’s value in what a model has learned about a product before it retrieves a page. A paid placement doesn’t directly replace that. Being discussed consistently in relevant sources may help, although I wouldn’t describe any organic position as permanent protection.

Josh: Agents are moving from recommending software to buying it. When an agent has to fulfill "make me a board deck by Monday" and picks between Gamma and someone else, it's not reading your homepage. It's evaluating structured capability, pricing clarity, integration surface, and whether it can complete the job without a human. What has to be true for Gamma to win that call, and how much of your 2027 roadmap is being legible to an agent versus delightful to a human? This is the question I think decides the next three years.

Ravish: The agent needs to know Gamma exists, understand what it can do and what it costs, and be able to complete the task. Clear feature pages, readable pricing and a reliable connector all matter. We’re seeing more attention to product and pricing pages alongside listicles.

We’ve had cases where a model recommends Gamma but doesn’t call the integration. Getting recommended and getting used need separate work. I expect the second to matter more as assistants take on more of the task.

Our roadmap, like most, has largely been designed around a person using the product on a screen. We also need to consider an agent reading the product information and trying the integration. I don’t have a percentage to put on that split, but it needs to be part of planning.

Josh: Headline call. End of 2027: is answer share still winnable from a standing start, or has it hardened into five names per category like Google page one did? Name the channel in Gamma's mix today that falls under 10% of signups by then, and the one under 10% now that ends up top three. No hedging. I'll put my calls next to yours in the piece.

Ravish: I expect broad ‘best tool’ questions to become harder to break into as the same names appear across the sources models use. More specific questions should remain open longer: a particular job, an industry or a language. We found it easier to get cited in some non-English markets where there was less competition.

The opportunity I’m watching is an agent completing a task with Gamma instead of sending someone to sign up. Our channel reporting doesn’t yet capture that as a distinct acquisition path.

If I were starting an AEO program this week, I'd take three things from Ravish's answers.

Thanks to Ravish for answering in detail, including on the parts that are still open at Gamma.