A Google VP just spelled out what AI search rewards: first-hand content that nobody else can publish. If your blog runs on AI-written posts, you're likely missing from AI answers. What gets you cited is the knowledge your team already has, starting with your product docs.

Open your latest blog post and ask one question: what in here could only your team have written? If the honest answer is nothing, you've published commodity content: common knowledge anyone could write, with nothing new for the reader.

I build content engines and train AI agents to write for companies, so I see this from the inside. Generative AI made this kind of content almost free to produce, and that's exactly why it stopped working. 

When every tech company can publish the same tips article in 30 seconds, no version stands out, and AI search has no reason to pick yours.

How to become the brand AI search cites

Whenever tech founders ask me how to win at SEO, GEO or whatever acronym comes next, I give the same answer: become the most helpful brand on the internet for your ICP (ideal customer profile).

Everyone usually expects a trick. But that answer is the whole strategy.

What Google calls noncommodity content

Google now says it too. In June, Brendon Kraham, Google's VP of Search and Commerce for Global Ads Solutions, published Good SEO is good GEO on Think with Google. His advice to marketers fits in 6 words: “You win by being relentlessly helpful.” His example is a running store that skips the generic "Top 10 Things to Consider" article and creates content about why one customer's shoe collapsed. He calls that noncommodity content. Google's guide to optimizing for generative AI features uses “7 Tips for First-Time Homebuyers” as its example of commodity content and tells site owners not to publish what an AI model could easily produce.

Why writing for your ICP matters even more in AI search

Knowing exactly who you're writing for matters even more in AI search, because the assistant increasingly knows who's asking too. ChatGPT's memory now references past conversations to give answers that feel “more relevant and tailored to you.” 

In January, Google added Personal Intelligence to AI Mode, an opt-in feature that uses context from Gmail and Photos to deliver “tailored responses in Search, just for you.” People also share more about themselves when they ask: Google says the average AI Mode search is triple the length of a traditional query.

So when a founder of a 10-person B2B startup asks an assistant for help, the model often already knows their role, company size and what they've asked before. My bet: the page written for exactly that founder has a far better shot at making it into the answer than a page written for everyone.

Why AI skips generic content

The case against generic content is simple. If a model can write your article from what it already knows, it can answer your reader's question without your article. Your post adds nothing new, so the answer has no reason to link to you. Mass-producing pages to game AI answers also breaks Google's scaled content abuse policy. A tech company publishing 40 AI-written posts a month is paying for content search engines have asked it not to publish.

For a decade, content strategy was a volume game. In AI search, one piece of proprietary data, like a benchmark your team actually ran, carries more weight than a month of generated posts.

Why proof and first-party data beat size

The encouraging part: real evidence gives small companies an edge. The researchers who coined the term GEO tested this in a paper accepted to KDD 2024. When they added sources, quotes and statistics to a page, it became up to 40% more visible in AI answers. Smaller sites that ranked lower in regular search got the biggest boost.

Your first-party data counts as that kind of proof: your own numbers, customer results and even failed experiments. When you publish them, you add new context to the web that no AI model could have produced on its own.

How do large language models synthesize your brand narrative?

When someone asks ChatGPT, Perplexity or Google's AI Mode about your company, the answer comes from many pages at once.

Most AI assistants follow the same pattern. First, they retrieve relevant pages from a search index and write an answer from them, with links to the sources. Google uses its own index, and Vercel notes that ChatGPT, Copilot and Meta AI use Bing's. Second, many use query fan-out: the model runs several related searches at the same time to fill in gaps. Google's guide gives the example of a question about a weedy lawn turning into separate searches about herbicides, weeding without chemicals and prevention.

Apply that to your company. "Is [your company] good for SOC 2 prep?" might pull from your pricing page, a comparison with a competitor, a Reddit thread, a review and your docs. The answer combines whichever pages come up in those searches.

Most of those pages aren't yours. University of Toronto researchers ran tests in 2025 comparing Google with AI assistants. For U.S. software product searches, Google's top results split almost evenly between company websites (43.7%) and third-party sources like reviews and independent publications (45.4%). GPT with web search used third-party sources for 72.7% of its citations and company websites for 26.7%. 

The AI tools also agreed with each other less about lesser-known brands than about famous ones, which is where most tech startups sit. (The paper hasn't been peer reviewed yet and tested 2025 model versions, so treat the exact numbers as a rough guide.)

This is the part founders underestimate most. AI answers about your company are shaped mostly by what other people write, plus whatever facts you make easy to find.

Founders spend months refining their homepage copy. Meanwhile, AI assistants form their view of the product from documentation, changelogs and customer conversations the founder rarely reviews.

That gives you 2 jobs. Give customers, reviewers and communities something real to talk about, and make your own facts clear and easy to check. That second job is where your engineers come in.

Why your engineers must write technical documentation for LLM evaluation

Buyers now ask AI assistants to compare products before visiting your site, and developers let coding agents choose tools for them. Your documentation is what the model reads to make that call.

Vercel noticed early. In June 2025, the company reported that ChatGPT sent around 10% of its new signups, up from 1% six months earlier. Its content team checks every page with one question: “Could a competitor easily replicate this tomorrow?” If yes, they go deeper.

Content written by engineers usually passes that test. Tailscale's 2020 post How NAT traversal works, by engineer David Anderson, explains how the product connects devices through firewalls. It includes real problems only their team would know, like a university guest Wi-Fi network that blocked the traffic their product relies on. No AI model could have written it first.

The most valuable content a tech startup owns already lives in its team. The companies that win AI search will be the ones that get that knowledge out of people's heads and onto the page.

5 things I'd ask every tech startup to do

1. Write the pages only your team can write, and make them expert-led. Document your limits, benchmarks, known issues and when a customer shouldn't choose you. The Toronto researchers found people use AI to build shortlists with reasons, so honest trade-offs give the model a reason to include you. Publish these pages under your experts' names, since their first-hand insight is exactly the context AI can't produce on its own. The hard part is getting your internal team and busy experts to write, and tools like Utopica.ai can help you turn their expertise into published content.
   
   
   

2. Publish docs as plain HTML, with a Markdown copy. Vercel found that most AI crawlers don't run JavaScript, so docs that only load in the browser can look empty to them. The llms.txt proposal also suggests offering a Markdown version of each page at the same URL with .md added.
   
   
   

3. Add an llms.txt file for AI agents. Google Search ignores it, but coding agents use it. OpenAI, Anthropic and Google's Gemini team all publish one for their developer docs, and Chrome's Lighthouse now checks for it.
   
   
   

4. Keep one dated source for every fact. Give pricing, rate limits, integrations and version numbers one page each, plus a dated changelog. Vercel reviews its content at 30, 90 and 180 days. When your site lists conflicting numbers, AI answers can end up quoting your old pricing.
   
   
   

5. Test how AI tools describe your product every month. Do this yourself, no matter what AI visibility tools advertise. Google states that no third-party tool has access to its internal ranking or AI systems. Write 20 questions a real buyer would ask ("best tool for X," "you vs. your competitor," "does it support SSO?") and run them in ChatGPT, Claude, Gemini, Perplexity and Google's AI Mode. Use a clean session each time: logged out in a private browser window where the tool allows it, or a temporary or incognito chat with memory turned off. These assistants remember you, and you've likely asked them about your product many times, so a logged-in test reflects your own history instead of what a buyer sees. For each answer, check whether you're mentioned, whether the facts are right and which sources are cited. Turn every wrong answer into a docs fix. For Google specifically, Search Console's new Generative AI performance report shows how often your pages appear in AI features.

Then judge the results by leads, sales and signups. No outside tool can see inside these AI systems, so your own funnel is the real test. That's how Vercel knew ChatGPT mattered.

My bet for the next few years

AI made generic content free, and that's good news for tech companies. When anyone can publish the same article in 30 seconds, the companies that stand out are the ones with something real to say. Tech companies have plenty of that.

You know your customers better than any model does. Your team has the numbers and the answers buyers keep asking for. Write for your ICP, back it up with your own data, turn your engineers' knowledge into clear docs and check what AI says about you every month.

Keep doing that, and you become the most helpful brand on the internet for the people you serve. That's the brand people trust, and the one AI search has every reason to cite.