How to get AI to recommend your product

How to win the AEO/GEO game on a startup budget.

About Kevin

Airbnb, Shopify, Reddit, and G2 have billions riding on SEO & AI search… and they all use the same advisor – Kevin Indig. He publishes original research through his newsletter, Growth Memo, and advises companies whose entire business model depends on getting discovered. Watch the full interview on YouTube

Can a small company outrank the big guys?

Q: Does AI search give startups an opportunity to leapfrog incumbents?

Yes  –  but only in a niche, never head-on. One of the biggest ingredients in AI visibility is authority: "how known are you, how good is your reputation, and by how many people?" Big companies win on authority by default. But AI answers are specific to the question asked, and that's where niches become powerful.

"A startup cannot be successful if they tackle Salesforce head-on... There's a famous saying: if you get 1% market share of Salesforce, you have a unicorn."

In plain terms: nobody can out-rank Salesforce for "best CRM." But for "best CRM for law firms" or "CRM for agencies with client-approved workflows"  –  Kevin's examples  –  a focused startup can absolutely become the answer AI gives. Because you're only competing in that narrow slice, "you have a much higher chance to create stronger awareness, create better content, and position yourself accordingly." Win the slice, then expand to the next one.

Kevin sees this working in reverse too: a large task-management client of his is now pushing into industry verticals (like construction) and finding "established startups that already are pretty good in those verticals" blocking the way. The niche defense is real.

What's the best content to create?

Q: What's the single best way to get cited by AI?

Publish original data that nobody else has. Remember the training cutoff: anything the AI doesn't already "know," it must find through live search. If you're the only source of a piece of information, you're the citation.

"LLMs crave structured data  –  numbers, lists, reports, benchmarks, case studies  –  and they crave net new data that helps them give a better answer."

Two conditions apply:

1. The data must be about your topic. A CRM company publishing an employee-retention study might get press, "but that does not help LLMs recommend you when it comes to CRM." Publish a CRM benchmark instead.

2. Answer real questions directly. Pair your data with plain, word-for-word answers to the questions people actually ask. "That is the one-two punch combo that gets you cited, and that builds strong defenses, because it's very hard to displace you."

Q: Re. topics, how do you find the questions worth answering?

Your customers have already told you  –  it's sitting in your call recordings. "Take 30, 40, 50 transcripts of sales conversations" from tools like Gong or Chorus, split them into won and lost deals, and interrogate them with AI: What problems do these customers face? What do they use today? What alternatives did they consider?

"We did this at Ramp... and we just created blog content that performed incredibly well, because we knew that our customers think about these questions, and we have good answers."

Review mining also works – your own reviews and competitor reviews often contain answers to people’s implicit questions, so you can reverse-engineer the review content to generate the questions. (Matt’s addition)

The zero-cost starter version: Google your own brand and answer every question in the "People also ask" box.

Q: What about ordinary informational content  –  guides, definitions, how-tos?

That era is over. "Creating informational content does not really make sense anymore." When someone asks a basic question, the AI simply answers it  –  studies show almost nobody clicks through to a website afterward. If you need glossary-style basics on your site, "do it with AI. Don't invest too much effort into this."

What still earns attention is content that's hard to replicate:

• Primary research (your own data, as above)

• First-hand experience  –  "Instead of 'here are the 10 best CRMs,' write 'I spent 100 hours stress testing 3 CRMs, and here's a detailed breakdown.'"

• Thought leadership  –  genuinely new ideas your audience hasn't considered

Kevin calls these "the three macronutrients" of a healthy content diet. The catch: unlike old-school SEO, where search volume told you exactly what to write, this content is "much harder to forecast. It's less comfortable for most companies, and that's why they shy away  –  but sticking to those old guns is usually what gets you into trouble."

Q: How good does the content need to be?

World-class  –  literally. Because AI has made content creation nearly free, the volume game is dead. The question to ask of every piece: "What would it take for this to be the best piece of content about that topic on the web? Pretty intimidating, but that's the standard." For a small team, that means fewer pieces, each one genuinely the best answer to a narrow question.

How to create and structure content for AI

Q: Should content be short?

Concise, which isn't the same thing. "The piece can be 3,000 words long and still be concise. Cut out the fluff, cut out the lore... every sentence can stand on its own, no passive wording."

Structure matters enormously, because of two findings from Kevin's analysis of over a million AI answers:

1. AI reads the top of the page hardest. Models favor the first 20–30% of your content. So put a summary or key-takeaways box at the very top, and write like a journalist: most important facts first. (If you know the Minto Pyramid  –  conclusion first, then support  –  that's the shape.)

2. AI extracts tiny amounts. Most Google AI services pull only about "250 to 300 tokens" from a page  –  roughly 200 words. The AI isn't reading your essay; it's hunting for the one crisp paragraph that answers the question. Make that paragraph easy to find.

Q: So should each question get its own short page, rather than one big guide?

Mostly yes. For scaled content, short and pointed wins  –  "most brands write too much about a topic. Most content can be a lot shorter." One exception: every company should keep a single FAQ page covering the basics. "Google your own brand name, see what Google shows in terms of “People also ask” questions, and just answer these questions on one single page."

Q: Google always said "write for humans." Is it finally true, or is there a new set of tricks?

Both. Human behavior still drives classic search rankings, and "classic search is still the skeleton of the body that is AI search." But  –  a striking stat  –  Cloudflare reports that bots now send more web requests than humans do.

Kevin's rule of thumb for executives: "Write it so that a human would enjoy it, but do not be Malcolm Gladwell or Mark Twain. This is not a story. At the end of the day, if the bots don't like the content, the humans won't see it." (His own Substack is the exception that proves the rule: audience-building writing, where people come to you directly, plays by different rules than content designed to be found.)

Q: Good brand marketing is emotional and opinionated. Will that hurt your AI visibility, since models seem to prefer dry, balanced content?

Not if you keep the two separate. "When it comes to the data, you need to be dry like a scientist. When it comes to the storytelling, you can be a full-blown marketer."

Present facts the way Wikipedia or a research paper would  –  clear methodology, neutral tone, stated caveats. (Wikipedia gets cited constantly, partly because it's objective.) Then bring all the drama you want to the narrative around it. "Be the scientist that pairs up with a strong marketer."

Where to publish and distribute

Q: Which platforms actually get read and cited by AI?

More than you'd think  –  and YouTube tops the list. "YouTube is probably the most important platform now when it comes to Google's AI properties" (AI Overviews, AI Mode, Gemini  –  and increasingly ChatGPT too). It's "incredibly often cited and correlates highest with brand mentions." Bonus: video is harder for competitors to copy and builds more human trust.

The broader principle is old-fashioned repurposing: one strong piece becomes a blog post, a newsletter, a YouTube video, and clips on social. LinkedIn gets cited heavily for B2B topics; Reddit more for consumer.

And crucially  –  don't lock your content away. "The worst thing that can happen is you put all that effort into creating an awesome piece of content, and then LLMs cannot get it because it's hidden behind some paywall or gate." That gated PDF whitepaper is invisible to AI. Publish an open version somewhere.

Q: For YouTube specifically  –  what matters? Shorts vs. long form?

The transcript is everything. "Transcript is what LLMs read to understand a video." Long form (7–10 minutes) has better odds, but cutting a webinar into question-and-answer clips works well too. The AI reads the transcript, chapters, title, and description  –  align them all.

The best-in-class move sounds almost comically literal: know the exact questions (prompts) you want to show up for, and "have the person in the video say the prompt verbatim so it lands in the transcript." For B2B, the channel itself doesn't matter  –  only the video.

Q: Do post comments, podcasts, and hosting location matter?

• Comments: Yes  –  "I've seen comments getting cited." Write a substantive comment (not one line), and it helps to be upvoted high on the page, since AI reads the top of pages hardest.

• Podcasts: Indirectly. Kevin hasn't seen podcast platforms cited, so "upload the podcast to your website and show the transcript below it." For niche topics with little existing content, an expert conversation transcript can win citations.

• Authoritative hosts: If your content appears on a well-known site (say, your investor's), AI pays more attention. "The more authoritative the site, the more willing LLMs are to invest retrieval budget"  –  meaning they'll crawl more of its pages. Unknown sites might only get a few pages looked at. You can see how often AI bots visit you in your server logs or CDN dashboard (Cloudflare, Akamai)  –  worth checking, because great content that never gets crawled is invisible.

Building reputation – Reviews, Reddit, and PR

Q: Do review sites like Google Maps, G2, Capterra, and Trustpilot matter?

Enormously. Kevin, who has worked inside G2's data: "Reviews are the trust layer of LLMs." Because reviews are written by verified users and are relatively objective, AI leans on them heavily to decide "how you're being recommended, in what order, and for what person."

Practical moves:

• Run review campaigns. Ask new users for a review after onboarding or a positive NPS response.

• Bias them fairly (Matt's addition, which Kevin endorsed "100%"): after an NPS survey, send promoters to the review site and detractors to a feedback flow.

• Don't fear imperfection. All-five-stars looks fake. "Even sometimes a three-star review, if it's balanced and fair, can actually increase trust."

• Mine reviews  –  yours and competitors'  –  for positioning insight, product feedback, and sentiment risks.

Q: Reddit is famously heavily weighted in AI training. Can you influence it?

It matters a lot, and it's dangerous to game. Reddit feeds both training data and live citations. But: "Reddit has very strong antibodies  –  thousands, maybe millions of mini-communities that all hate to be manipulated." Obvious brand promotion gets deleted or publicly mocked.

Kevin's honest framing: this is a 6-month-plus investment with no guaranteed payoff. Lurk first, learn the norms, contribute genuinely, and only later become brand-forward. Some communities will never welcome you; some brands are simply "burned" there already.

His cheap validation trick: "Advertising on Reddit is a great way to figure out if your audience is there and what they think about you." Spend a little on ads first; the engagement (or hostility) tells you whether deeper investment makes sense.

Q: How does PR help  –  is it still about backlinks?

No  –  it's about mentions. "Mentions are the most important thing." Backlinks still correlate with visibility in studies, but Kevin believes that's "much more because of the mention on a strong authoritative site, and less because it's actually linked." The AI doesn't need to follow a link; it just needs to see your name in a trusted place.

A surprising loophole: AI doesn't distinguish paid from earned coverage. "It still very much looks like LLMs do not differentiate between sponsored content and organic content." So syndication, advertorials, and even plain advertising on strong publishers all count. Add influencer collaborations (Substackers, bloggers, LinkedIn voices) and, if you have well-connected investors, their distribution.

Q: Founder brand or company brand?

They're inseparable  –  invest in the founder. "People follow people." His examples: Alex Karp, who rarely discusses Palantir directly but constantly ties world events back to it; and Stripe's CEO launching a video podcast  –  "that should tell you everything about how important that is."

The playbook: pick your strongest ambassador (usually the founder), give them a narrative  –  the founder's arc, the problem, why it matters  –  and put them on LinkedIn and YouTube. For AI specifically, founders get the most exposure in major media, so "give them scripts  –  keep these three sentences in mind and say them" so the key positioning language lands in high-authority coverage.

Measurement  –  how do you know it's working?

Q: The old SEO funnel was search → click → convert. What replaces it?

Nobody clicks anymore, so stop measuring clicks. "If you're measuring traffic as an indicator of whether your AI strategy works, you're in a really bad place." We're heading toward a near-zero-click world  –  a leaked ChatGPT figure showed a 0.69% click-through rate on a citation with around a million impressions.

What actually happens: someone gets a recommendation from AI, then "opens a new tab, goes directly to the site, and signs up." The visit looks like it came from nowhere.

So the single most valuable measurement is the humble "How did you hear about us?" field, with an AI/ChatGPT option. This is called self-reported attribution, and it's not a rounding error: "One of my clients right now is seeing 10% of net new clients coming from AI."

Q: What's the full measurement framework?

Three layers:

1. Leading indicators  –  early signals you're on track: how often AI mentions you, cites you, your share of voice versus competitors, plus classic search rankings (which still feed AI answers).

2. Guardrails  –  how you show up, not just whether. Being mentioned as the cautionary tale is worse than not being mentioned. (His example: LastPass appearing in answers as the security-breach warning.)

3. Lagging indicators  –  the business results: self-reported attribution, customers mentioning ChatGPT in sales calls, pipeline, revenue.

Matt's summary, which Kevin confirmed: it takes some faith. If people ask the questions, you're the right answer, and you're getting mentioned  –  trust that quality traffic follows, sanity-check with conversion rates and self-report, and don't demand click-level proof that no longer exists.

Q: Are the new AI-visibility tracking tools worth buying?

Judge them by the "action layer," not the tracking. "Everybody and their grandmother starts a prompt tracker"  –  tools that show where you appear in AI answers are now commodity. The differentiator is whether the tool helps you act: spot a topic where a competitor is winning, then refresh old content and draft new pieces  –  "with the strong caveat that you want a human in the loop."

Be suspicious of tools promising to automate distribution (getting you mentioned places): "We've seen that in the SEO space, and it ended poorly." Names mentioned: AirOps (his client  –  disclosed), Profound, Scrunch, plus Semrush and Ahrefs adding AI features.

Q: Is it risky to use AI to create the content itself?

Risky if AI is the author, low-risk if it’s the assistant. "If you go to Claude and just say 'write me an article about X,' that is the wrong approach." Let AI do the legwork  –  research, drafts, stress-testing, graphics  –  then add what only you can: experience, primary data, a real quote from the CEO. "The best content still comes from humans plus machine." Routine refreshes of existing content, though, can be automated with good results.

Q: Is schema markup still worth doing?

(Schema = invisible code labels that tell machines "this is a price," "this is a review score," etc.) Bottom of the priority list. Some studies, including his own, show a modest effect, so "if it's easy for you to implement, go for it"  –  but crawlability, content, and distribution all matter far more.

The cheap, immediate wins for a lean startup

Q: For a startup with a limited budget, what are the highest-leverage moves?

Kevin's priority order:

1. Make sure AI can actually read your site. This is table stakes  –  "that's the minimum viable scope." AI agents "get stuck more on websites than people might think." The free test: ask ChatGPT and Claude for specific information that's on your site, and see whether they find it and whether they cite you or a competitor. His new research finds pricing is the hardest information for AI to retrieve because companies hide it  –  so if you're attacking an incumbent, "just be transparent about your pricing." An instant edge.

2. Answer questions nobody has answered well. Write out 10–20 questions (prompts) people would ask about your category, check what AI currently says, and create better content where the answers are weak.

3. Then think about PR  –  in the new-media sense: original research, mentions in trusted relevant Substacks and blogs, influencer collaborations, not just traditional press.

"Most companies are still in level one." Just doing these three puts you ahead.

The big picture  –  future of the marketing organization

Q: Teams built entire departments around SEO and are watching the traffic evaporate. What does the marketing org of the future look like?

Corporate journalism and audience building. His anchor case: HubSpot  –  the company that invented "inbound marketing"  –  lost roughly 80% of its organic traffic, "and the reason that loss does not correlate with their stock price is because much earlier, they invested in audience building." They bought The Hustle and built "an army of newsletters, podcasts, YouTube channels." Okta hired six full-time journalists into an internal newsroom.

Structurally, the old silos  –  SEO, brand, performance, product marketing  –  "are melting together." The best teams he works with are "single-threaded, composed of channel experts, that go after a singular goal." And that's the startup's built-in advantage: "A Salesforce is not going to create such a team. But a new attacker can pull all these strings  –  be present on all these outlets with a unified strategy, unified messaging  –  and attack the market with a united front."

Matt's closing synthesis, which Kevin endorsed with a one-word "Bingo": building an owned audience  –  newsletter, social, community  –  is never wasted effort. Do it well and it does double duty: it reaches people directly and it's exactly what makes AI recommend you.