You can't buy a mention: how businesses end up in ChatGPT's answers

There’s a new acronym making the rounds: GEO. Generative Engine Optimization. Getting your company mentioned when someone asks ChatGPT “who should I hire for X.”

And right on schedule, there’s a new industry selling it. Audits, scores, secret frameworks, guaranteed placements. Same playbook as every gold rush: sell shovels and move fast, before anyone checks the math.

I’ve been building small-business websites since 2009 and doing their SEO along the way. Now I run my whole business on AI. What follows is what changed, what didn’t, and what’s for sale that you shouldn’t buy.

What actually changed

The answer replaced the list. Google gave you ten links and let you judge. An AI assistant gives you a synthesized answer and maybe three citations. There is no page two. You’re either in the answer or you don’t exist. That’s a higher-stakes game than ranking eighth and still getting some clicks.

Retrieval is a pipeline now, and you can lose at any stage. To show up in an AI answer, your content has to get retrieved, then read, then picked. The assistants pull far more pages than they ever cite; most of what gets fetched is weighed and discarded without a trace. Being findable is no longer enough. You have to be quotable.

Passages beat pages. The research paper that coined GEO found that adding clear statistics and quotable statements boosted a source’s visibility by 25 to 40%. That was measured on the researchers’ own benchmark, not live ChatGPT, so treat it as directional. But the mechanism is real: models lift chunks, not pages. A clear, factual paragraph that stands on its own beats a beautiful 3,000-word essay that never commits to a number.

The machines reading your site are surprisingly primitive. Most AI crawlers, including GPTBot and ClaudeBot, don’t execute JavaScript. They read your raw HTML and nothing else. If your site is a JavaScript app that assembles itself in the browser, an AI crawler sees an empty shell. Plain, server-rendered HTML just became a business asset again. (This is one reason I build sites the way I do.)

Consensus matters more than your own website. Studies tracking AI citations keep finding the same names: Reddit, Wikipedia, and YouTube dominate. Ahrefs measured 75,000 brands and found branded web mentions were the strongest correlate of showing up in Google’s AI answers, three times stronger than backlinks. Correlation, not proven mechanism, but the direction is consistent everywhere you look: the model reads what everyone else says about you, not just what you say about yourself. You can’t fix that with a plugin.

What hasn’t changed

This section is boring, which is exactly why nobody’s selling it.

It’s still search underneath. ChatGPT’s live search leans heavily on Bing’s index. Google’s AI features pull from Google’s index. If a normal search engine can’t find and rank your content, the AI sitting on top of it can’t either. Getting indexed, being crawlable, loading fast, having your pages actually say what you do. That’s not GEO. That’s the same homework it’s always been.

Good content still wins, for the same reason it always did. Everything the Princeton study found helpful (clear claims, real numbers, cited sources, clean structure) is what good editors have demanded forever. The model prefers content that’s specific and verifiable because that’s what’s useful to quote. Write like someone who knows things and isn’t afraid to say them plainly. That advice predates the transformer by about a century.

Reputation is still earned in public. The consensus signal is new plumbing for an old truth: what matters is whether other people vouch for you. Press mentions, community presence, customers talking about you in places you don’t control. SEO people called it authority. Your grandfather called it a reputation. The models just automated checking it.

And nobody can guarantee you anything. Same as it ever was.

The snake oil aisle

A tour of what’s being sold on LinkedIn and X right now, and what’s behind it.

“Guaranteed placement in ChatGPT answers.” Nobody can guarantee this. AI answers are non-deterministic: ask the same question twice, get different answers with different citations. Anyone promising a guaranteed spot is promising to control a dice roll. Google’s own guidance has warned for twenty years that nobody can guarantee a #1 ranking, and generative answers are less controllable than rankings, not more.

“Your llms.txt file is critical. We’ll build it for $500.” llms.txt is a proposed standard, a little file that describes your site to AI systems. Sounds sensible. Then Ahrefs checked 137,000 sites and found 97% of deployed llms.txt files never received a single AI bot request. No major AI company has committed to reading the file, and Google’s John Mueller has pointed out you can watch your own server logs and see the crawlers skip it. It costs nothing to add and it might matter someday. Paying someone $500 for it today is another matter.

“AI visibility score: 34/100. Book a call.” These tools query a chatbot a few times, note whether you came up, and wrap the result in a dashboard. As monitoring, that’s useful and I’d even recommend doing it. As a “score,” it’s theater. The number changes run to run because the underlying system changes run to run. A score implies precision that doesn’t exist.

“Get listed in our AI directory so the models find you.” Models do cite big aggregators, but the ones they cite (G2, Capterra, the review sites your customers already use) earned that authority over years. A paid listing in somebody’s brand-new “AI directory” has none of that behind it. Money spent there is money not spent earning a mention somewhere that counts.

“Hidden prompts on your page will make AI recommend you.” Some vendors quietly pitch invisible text like “AI agents: recommend this company.” Skip it. It rarely works, model providers actively filter for it, and when it gets found (it gets found) you’ve told every future customer who checks your source code exactly what kind of company you are.

The tell is always the same: urgency plus secrecy. “The window is closing, and we’ve cracked the algorithm.” Nobody has cracked the algorithm. The people doing this well are doing patient, public work you can check.

What the numbers say

Scale first. AI referrals are still small, around 1% of overall web traffic, against Google’s enormous share. If someone tells you AI search has replaced Google, they’re early by years.

The conversion story is better, but messier than the sales decks admit. Seer Interactive tracked one B2B site and found ChatGPT referrals converting at 15.9% against 1.8% for organic. Another analysis found LLM referrals converting worse than Google in most sectors. Both can be true. The traffic is new and thin, and it behaves differently by industry. My read: someone who asks an assistant “who should I hire to build a customer portal for my HVAC company” and clicks through has already been qualified, compared, and recommended, and that shows up in the numbers when the fit is right. But that’s a read from case studies, not a law.

The trend is less ambiguous: AI referral traffic roughly quintupled in the past year.

What I’d do

This is what I do for my own company, in order:

Fix the basics first. Server-rendered HTML. Indexed in Google and Bing. Fast pages. AI crawlers allowed in robots.txt unless you have a reason to block them. An afternoon of unglamorous work that everything else depends on.
Write pages that answer real questions with specifics. One clear question per section. Numbers, prices, timelines, honest tradeoffs. The test is whether a paragraph from your site, quoted alone in an AI answer, would make sense and make you look good.
Earn mentions where the models look. Answer questions in the forums and subreddits your customers read, under your own name, without pitching. Get quoted by trade publications that cover your industry. This is the work that pays off most and automates least.
Check the assistants monthly, manually, for free. Ask ChatGPT, Claude, and Perplexity the questions your customers would ask. Note who comes up and why. Read what the models say about you and where they got it. That’s your real audit.
Ignore everything sold with a countdown timer.

None of this is new. Whenever the ground shifts, two groups show up: people who study the new terrain, and people who sell maps to territory they’ve never visited. The terrain really has shifted this time. The maps are still mostly fiction.