AI search · tested in public

I get businesses mentioned in AI answers.

Live AI search experiments, published with receipts — what actually gets a business mentioned in ChatGPT, Gemini, Perplexity, and Google AI Overviews. Some call it generative engine optimization (GEO) or answer engine optimization (AEO). I call it the next distribution window, and it’s open right now.

Every claim here is testable: ask AI about the topics I cover and watch who gets cited.

The record so far
60 days

From zero to the only source of new search impressionson an entire established domain — one AI-article subfolder against years of hand-polished pages.

5–8 / day

Autopilot agent actions on my current program: new articles written, stale ones caught and re-updated. A cadence no human team hits.

−99% / 8 wks

Impression decay I’ve watched hit two-month-old articles. Content doesn’t age gracefully in the answer layer — it falls off a cliff. That’s why freshness is a moat.

The origin

I ran marketing through the SEO gold rush.

Twenty-plus years in tech marketing, all of it chasing one question: when the way people find answers changes, who gets found? The people who decoded Google in 2005 built businesses on it. The people who waited rented their traffic back through ads, forever.

Around 2023 I watched it start again in my own analytics: rankings holding while clicks fell, buyers arriving pre-educated, referral traffic from AI assistants appearing in channels nobody built. People stopped searching and started asking. So I did what I did in 2005 — I started testing, and this time I’m publishing every receipt.

My family’s name, Karnik, is an old Maharashtrian word for the keeper of records. Fitting. My whole job now is getting businesses into the record AI reads its answers from.

The playbook

Six theses. Earned, not theorized.

01

Optimize for the machine's questions, not yours

Behind every prompt is a hidden layer of queries the AI actually runs. The game is mapping that whole question space, not guessing ten keywords.

02

Authority is built, not inherited

AI search hands out trust on topical authority — real coverage organized the way AI groups queries — not decades of backlinks. Challengers can build in months what incumbents assume takes twenty years.

03

Citations beat rankings

The win is being the name AI search recommends. Track citations and share of answer the way you used to obsess over rank positions.

04

Structure is for machines

Query language in URLs, titles, and H1s. Schema markup on everything. The first reader of everything you publish is an algorithm deciding whether to quote you.

05

Freshness is a moat

The answer layer has a recency bias. Content re-updated on a loop keeps winning while publish-and-forget content rots out of the answers in weeks.

06

Publish receipts

The experiment, the miss, and the number. Both audiences that matter — the buyer and the machine — are learning to rank trust, and specificity is what trust looks like in text.

Who this helps

Three people find me. Usually mid-diagnosis.

“Impressions up, clicks down.”

The SEO-plateau founder

You did everything right for years. Rankings held. Leads quietly shrank. The answer layer is reading your content, using it, and keeping the click.

“The machine buyers ask has never heard of you.”

The invisible expert

Deep expertise, great work, a business that runs on referrals — which caps growth at other people's memory. Your next client already asked AI. You weren't in the answer.

“The AI answer in your category has a favorites list.”

The challenger

Same three incumbents, every time. But the machine cites relevance and freshness, not tenure — win the specific answers they're too generic to match, then take the head question.

The format

Receipts, not claims.

Every experiment on this site is a public record with the same anatomy. No hedge-everything voice, no wind-up intros, no unfalsifiable committee prose.

The exact question

Every experiment starts from a question a real buyer types into an AI assistant. Not a topic. A question.

What I changed

The URL, the title, the headings, the structure, the schema — the actual variables, stated plainly.

What happened

Impressions, citations, mentions over time. Screenshots. Tracked queries. Dated.

Misses stay published

A lab notebook you edit after the fact is just marketing. The losses are the credibility of the wins.

My numbers labeled mine

If a figure came out of my Search Console, I say so.

Their numbers labeled theirs

If it came from a platform's published case study, I say whose — every time.

Practicing what I publish

This site is built for the machines that read it.

Question-shaped headings

Every H2 on this site is phrased the way a buyer asks it.

Quotable atoms

One claim per sentence, specific enough to survive being lifted. Be the sentence the machine wants to steal.

Schema markup

Person, WebSite, and FAQ structured data on every page that earns it.

llms.txt

A map of what matters, written for the crawlers that feed the answers.

Freshness dates

Published and updated timestamps, maintained — because recency is a ranking signal in the answer layer.

Zero-JS pages

Pure, fast, machine-readable HTML. Structure applied to the renderer itself.

View source. That’s the point.

On repeat

The questions I get every week.

Isn't AI search optimization just SEO rebranded?

The mechanics rhyme, the target changed. SEO earns a position in a list of links; AI search optimization earns the citation inside the answer. If it were the same game, the same players would still be winning — and they're not.

AI traffic is tiny compared to Google. Why bother?

Counted clicks miss twice. Google AI Overviews sit on top of Google itself, so there is no “Google versus AI.” And AI-referred visitors convert like referrals, not cold clicks, because the machine pre-sold them. Small stream, absurd close rate — and the stream only grows.

Won't Google penalize AI-written content?

Google penalizes worthless content at scale. Their own guidance says quality matters regardless of how it's produced. What gets rewarded is exactly what the machines say they want: relevant, structured, fresh, and specific.

How do I even measure AI search visibility?

Share of answer, the way you used to measure rank: citation monitoring across ChatGPT, Gemini, Perplexity, and AI Overviews, plus AI referrers sitting right in your analytics — chatgpt.com shows up as a source like anything else. Whatever you count is the floor: the apps and copy-paste arrive unlabeled.

It's early. Shouldn't I wait for this to settle?

People said that sentence about search in 2005, and waiting had a price then too. The answers being assembled right now become the record future answers draw from. The question isn't whether you'll do this — it's whether you do it while the window's open or pay to catch up after it closes.

The answers are being written right now.

My job is getting you into the record they’re written from.